<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Data by Toufik on Medium]]></title>
        <description><![CDATA[Stories by Data by Toufik on Medium]]></description>
        <link>https://medium.com/@databytoufik?source=rss-d41ecc85bfcc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*M_hRkr8JWo-ulY_k0YP4jw.png</url>
            <title>Stories by Data by Toufik on Medium</title>
            <link>https://medium.com/@databytoufik?source=rss-d41ecc85bfcc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 03:53:59 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@databytoufik/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How to Use Claude in Google Sheets (the New Add-on, Tested)]]></title>
            <link>https://medium.com/@databytoufik/claude-in-google-sheets-add-on-39102291ed2e?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/39102291ed2e</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[google-sheets]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 14:01:03 GMT</pubDate>
            <atom:updated>2026-11-12T14:01:03.167Z</atom:updated>
            <content:encoded><![CDATA[<h4>I gave it a messy 246-row sales export with five known problems and checked every answer. Here’s what it got right and where it falls short.</h4><figure><img alt="Before and after: a raw sales export with mixed dates, region spellings and a #N/A error, next to the cleaned tab where 240 of 240 orders match." src="https://cdn-images-1.medium.com/max/1024/1*hLWE2Ayd69oLeMV56JweXg.png" /></figure><p>To use Claude in Google Sheets, you install Anthropic’s Claude for Google Workspace add-on and open it from the Extensions menu. It opens a sidebar next to your sheet that reads the open file, writes formulas, builds tabs and charts, and, by default, asks before it changes anything. It’s in beta and needs a paid Claude plan.</p><p>I tested it once, on a made-up Q3 sales export with five planted problems, and checked its work against answers I had calculated separately. In that run (default model, Opus 5.5 at Medium effort) it got all four tasks right. It also made one mistake of its own along the way, then caught and fixed it. One run on one practice file isn’t a guarantee, so the last task below is how to check its numbers yourself.</p><h3>Which “Claude for Sheets” is this?</h3><p>Search results mix up several similarly named tools:</p><ul><li><strong>Claude for Google Workspace</strong> (this article): a sidebar add-on for Sheets, Docs and Slides, in public beta on all paid plans since <a href="https://claude.com/resources/articles/claude-now-works-in-google-docs-sheets-and-slides">October 6, 2026</a>. The sidebar itself is titled “Claude for Sheets”, which adds to the confusion.</li><li><strong>An older API-key add-on</strong>: puts formulas like =CLAUDE() in cells and bills an Anthropic API account. Many older tutorials are about this one.</li><li><strong>Claude Sheets</strong>: a simple spreadsheet inside the Claude app, in beta since November 10, 2026. It exports to .xlsx or CSV.</li><li><strong>Claude for Excel</strong>: the Excel version of the sidebar.</li></ul><h3>How to install Claude in Google Sheets</h3><p>You need a Pro, Max, Ultra, Team or Enterprise plan, and a Google account that’s allowed to install Marketplace apps. At work, your Google admin may need to install or allow it. One install covers Docs, Sheets and Slides.</p><p><strong>Step 1.</strong> In Google Sheets, open <strong>Extensions &gt; Add-ons &gt; Get add-ons</strong>, search for Claude, and pick the one by Anthropic.</p><figure><img alt="The Claude add-on by Anthropic in the Google Workspace Marketplace" src="https://cdn-images-1.medium.com/max/1024/1*UmgMxnfYcSpJPmQOnfga3A.png" /></figure><p><strong>Step 2.</strong> Click <strong>Install</strong> and allow the permissions. The list looks broad. According to Anthropic, the Docs, Sheets and Slides permissions only cover the file the add-on is opened in. It also asks to see your Google account email address and basic profile info.</p><figure><img alt="Google’s consent screen listing the six permissions the Claude add-on requests" src="https://cdn-images-1.medium.com/max/756/1*8NTtmDTKa922yQvC1PGuHg.png" /></figure><p><strong>Step 3.</strong> Open a native Google Sheet. If your file is an .xlsx you opened in Sheets, Claude won’t appear: use <strong>File &gt; Save as Google Sheets</strong> first. Files that were already open before the install need a reload.</p><p><strong>Step 4.</strong> Go to <strong>Extensions &gt; Claude &gt; Open Claude</strong>.</p><figure><img alt="Extensions menu with Claude &gt; Open Claude" src="https://cdn-images-1.medium.com/max/1024/1*l5XQd360_G7MaKGEzTcpUg.png" /></figure><p><strong>Step 5.</strong> Click <strong>Log in</strong>. The sidebar shows a short code and opens claude.ai in a new tab to confirm. Back in Sheets, it finishes connecting on its own.</p><p>On first run it shows a beta notice with a warning worth reading: files, emails and attachments can contain hidden instructions aimed at Claude, so only use it with content from sources you trust.</p><figure><img alt="Beta notice warning about hidden instructions in files" src="https://cdn-images-1.medium.com/max/291/1*H-Kj5lcLG69OWHyGYN-Rrg.png" /></figure><p>The sidebar has a model picker (mine defaulted to Opus 5.5 at Medium effort), built-in skills like /clean-data-sheet and /audit-sheet, file attachments, and a web search toggle that&#39;s off by default.</p><h3>The test</h3><p>The practice file is a made-up Q3 sales export: 246 rows, two tabs (Orders and a five-product Prices list), and five problems I planted on purpose:</p><ul><li>Six duplicate rows, as if the export ran twice for a few orders</li><li>Dates stored as text in three formats: 2026-07-03, 07/03/2026 and 3 Jul 2026</li><li>44 region cells with nonstandard spellings like “N.”, “EAST”, “East Region” and “west ”</li><li>44 prices stored as text, like “$249.00”</li><li>16 List Price cells showing #N/A because the product name has a trailing space</li></ul><p>Before testing, I calculated the correct answers with a separate script: 240 unique orders and $428,903.55 in revenue, plus totals by region and by month.</p><p>I sent four prompts in one conversation, using plain wording without hints about the planted problems.</p><figure><img alt="The messy export, with the sidebar scrolled back to the first prompt" src="https://cdn-images-1.medium.com/max/928/1*5lpadS7w1zKf2kcjxurV0A.png" /></figure><h4>1. Clean the export</h4><blockquote><em>This tab is a raw Q3 sales export. Make a cleaned copy in a new tab called Orders_clean: one row per order, real dates, consistent region names, numeric prices, and a Revenue column (Units × Unit Price). Leave the Orders tab unchanged. Then tell me what you changed and how many rows you removed.</em></blockquote><p>Claude works by running small scripts. In the default “Ask before edits” mode, each one shows a card with what it will do and how risky it is, and you choose Deny, Always allow or Allow once. There’s also an “Accept all edits” mode that skips the cards for ordinary edits. Every request was approved with Allow once: 11 approvals across the whole test.</p><figure><img alt="A permission card: Claude wants to run a script" src="https://cdn-images-1.medium.com/max/315/1*6l38PMMjK9F80rPqXEzz9Q.png" /></figure><p>It took about three minutes and four scripts. The new tab had 240 rows, and its reply listed the six duplicate order IDs it removed, the region spellings it merged, the 44 prices it converted, and the trailing spaces behind the #N/A errors.</p><figure><img alt="The cleaned tab and Claude’s summary of what it changed" src="https://cdn-images-1.medium.com/max/1024/1*ZCCMBt08Ng04gKec-EV9tA.png" /></figure><p>Its first pass put every date one day early, a timezone mismatch between its script and the sheet. Claude ran its own date check, caught it and rewrote the dates before replying. It read the slash dates as month/day, reasoning that it’s the only reading that keeps every order inside Q3.</p><figure><img alt="Claude explaining the date formats and the timezone fix" src="https://cdn-images-1.medium.com/max/420/1*W-qC3iPRZSK_EGyGkMJxjw.png" /></figure><p>That reasoning is sound for this file, but it wouldn’t settle everything on its own: nine of the slash dates land inside Q3 whichever way you read them. That’s why I didn’t rely on the totals. I compared the cleaned tab with my answer file order by order, and every one of the 240 orders matched on date, region, product, units, price and revenue. Claude was upfront about its own checking, too: it said it had checked the first and last rows and the date range, not every row in between.</p><h4>2. Fix the #N/A lookups</h4><blockquote><em>Some List Price cells in Orders show #N/A. Why, and how do I fix it without editing the Prices tab?</em></blockquote><p>It answered in under 20 seconds: the lookup needs an exact match, and “Desk Lamp ” witha trailing space doesn’t match “Desk Lamp”. The fix is to wrap the lookup value in TRIM:</p><pre>=VLOOKUP(TRIM(E2),Prices!$A$2:$B$6,2,FALSE)</pre><figure><img alt="Claude’s diagnosis of the #N/A errors" src="https://cdn-images-1.medium.com/max/420/1*CJ5Oa3AOyj40y74vs-kdTw.png" /></figure><p>It didn’t apply the fix on its own, because I had asked earlier to leave the Orders tab unchanged. It offered instead. When I said yes, it changed only the List Price column, then re-checked every row.</p><figure><img alt="Claude confirming the fix: none of the 246 rows show #N/A, and only column H changed" src="https://cdn-images-1.medium.com/max/420/1*6gLlRkk44NygmlA1nFgQMA.png" /></figure><p>I checked too: all 246 rows showed the right price, and the Prices tab was untouched.</p><h4>3. Build a summary and a chart</h4><blockquote><em>Add a Summary tab with Q3 revenue by region and by month, using formulas that reference Orders_clean, and a column chart of revenue by month.</em></blockquote><p>About two and a half minutes later there was a Summary tab with live SUMIFS formulas, a cell checking that both totals tie back to the cleaned data, and a labeled column chart. Every total matched my answers. Its reply ended with a fair caveat: it hadn’t looked at how the chart renders and suggested checking the axis labels. They were fine.</p><figure><img alt="The Summary tab with revenue by region and month, a column chart, and my check cells" src="https://cdn-images-1.medium.com/max/1024/1*9ijVWcxCbcsJtegOofu64A.png" /></figure><h4>4. Check its numbers</h4><blockquote><em>How did you calculate total revenue? Give me a formula I can put in a cell to check it.</em></blockquote><p>It explained the method (units times the selling price after discounts, not the list price) and gave a check that recalculates the total without using its Revenue column:</p><pre>=SUMPRODUCT(Orders_clean!F2:F241, Orders_clean!G2:G241)</pre><p>That’s a good start, but it still relies on the cleaned units and prices. These quick checks take a minute (they’re the “My checks” cells in the screenshot above):</p><ul><li>=COUNTUNIQUE(Orders!A2:A) and =COUNTA(Orders_clean!A2:A): the unique orders in the export should equal the rows in the clean tab (240 and 240). A gap means rows were dropped or duplicated.</li><li>The SUMPRODUCT formula above: the total should match the summary ($428,903.55).</li><li>=MIN(Orders_clean!B2:B) and =MAX(Orders_clean!B2:B), formatted as dates: they should sit inside the period you expect.</li></ul><p>They catch dropped rows, double counting and wildly wrong dates. They won’t catch a region mapped to the wrong name or a date that’s off but still in range. For that, pick five or six rows from the original export and compare them with the cleaned tab by hand, especially the messy ones.</p><h3>What Claude in Google Sheets can’t do</h3><p>From <a href="https://support.claude.com/en/articles/16951679-use-claude-in-google-docs-sheets-and-slides">Anthropic’s support page</a>, as of November 2026, while it’s in beta:</p><ul><li>One operation can run for up to six minutes before Google stops it, so split big jobs into stages.</li><li>It can’t read or edit Connected Sheets tabs (BigQuery). Extract the rows to a normal tab first.</li><li>It can’t create triggers, custom menus or scheduled refreshes.</li><li>It works only on the open file, and the sidebar has to stay open while it works.</li><li>Chrome, Edge and Safari are supported. In Firefox the sidebar loads but actions don’t complete.</li></ul><h3>Before you use it at work</h3><ul><li><strong>Your plan decides the terms.</strong> According to <a href="https://support.claude.com/en/articles/16951679-use-claude-in-google-docs-sheets-and-slides">Anthropic’s support page</a> (checked November 2026), consumer terms apply on Pro, Max and Ultra, and commercial terms with a data processing addendum apply on Team and Enterprise. Anthropic says add-on content isn’t used for training and is deleted within 30 days, with some exceptions, and that it doesn’t follow your organization’s custom retention settings. Zero Data Retention isn’t available, and it isn’t covered for health data under HIPAA. Beta terms can change, so check the page before you rely on them.</li><li><strong>Its edits carry your name.</strong> Google records them under your account in version history, like edits you made yourself.</li><li><strong>Connectors widen what it can see.</strong> Setup listed the connectors already linked to my Claude account (Google Drive, Gmail, Calendar). By default it only works on the open file, but connectors you turn on can let it reach other data. Check what’s connected before you open a work file.</li><li><strong>Hidden instructions are a real risk.</strong> Don’t point it at files or data from sources you don’t trust.</li><li><strong>Some files can’t be uploaded anywhere.</strong> For those, one option is to work on a synthetic copy with the same structure and no real values. I wrote about that in <a href="https://medium.com/@databytoufik/how-to-use-ai-on-company-data-without-uploading-it-0ae24ce3265d">How to Use AI on Company Data Without Uploading It</a>.</li></ul><p><em>I build data pipelines and automation for clients. More at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=claude_google_sheets"><em>databytoufik.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=39102291ed2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Use AI on Company Data Without Uploading It]]></title>
            <link>https://medium.com/@databytoufik/how-to-use-ai-on-company-data-without-uploading-it-0ae24ce3265d?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ae24ce3265d</guid>
            <category><![CDATA[synthetic-data]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[data-privacy]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 12:27:43 GMT</pubDate>
            <atom:updated>2026-11-12T14:06:30.908Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Three steps: make a synthetic copy on your computer, let the AI write the analysis from it, and run that analysis on the real file yourself.</em></p><figure><img alt="A real file with hidden values stays on your computer; a synthetic copy with generated names goes to the AI." src="https://cdn-images-1.medium.com/max/1024/1*70MOQ2VP04cXE8waiC_m1A.png" /></figure><p>AI can write the analysis code for a spreadsheet you aren’t allowed to upload. Give it a synthetic copy instead, with the same columns and made-up rows, and run the code it writes on the real file on your own computer.</p><p>I tested this on a practice file of 2,400 expense claims with made-up employee names, IDs and amounts. The script, the prompt and the practice file are on <a href="https://github.com/databytoufik/use-ai-on-company-data">GitHub</a>.</p><h3>Setup</h3><p>Install Python 3.13 from python.org, since SDV doesn’t support the newest version yet. On Windows, tick “Add python.exe to PATH”. On the GitHub page, click Code &gt; Download ZIP and unzip it to a folder that isn’t synced to OneDrive or iCloud. Then open PowerShell in the unzipped folder that contains make_synthetic_copy.py and run this once (the packages take about 1.3 GB):</p><pre>python -m venv .venv<br>.venv\Scripts\python -m pip install -r requirements.txt</pre><p>On a Mac, use Terminal, python3 in the first line and .venv/bin/python instead of .venv\Scripts\python. I tested on Windows only, and SDV won&#39;t install on older Intel Macs with Python 3.13.</p><h3>Step 1: Make the synthetic copy</h3><pre>.venv\Scripts\python make_synthetic_copy.py</pre><p>The script uses SDV, a library that learns a table’s patterns and generates new rows. For your own file, save it from Excel as CSV UTF-8, create a folder called private and put it there. Then edit the settings at the top of the script. The column names in them belong to the practice file, so replace each one with yours or remove it ([] for an empty list, {} for an empty TYPE_FIXES):</p><ul><li>REAL_FILE and SYNTHETIC_FILE: your file and the copy, both in private.</li><li>DROP_COLUMNS: columns your questions don’t need, such as free-text notes.</li><li>TYPE_FIXES: every column with people’s names set to name, and every ID column set to id with a pattern, such as X[0-9]{5} (an X and five digits). Pick a prefix your real IDs don’t use.</li><li>COPY_AS_IS: category columns and whole-number counts whose values are fine to share, such as departments or quantities. Never names, emails or IDs: their real values would be copied.</li><li>SENSITIVE: any other text column whose real values must not appear in the copy.</li><li>DATE_COLUMNS and DATE_FORMAT: every date column, and how the dates are written, such as &quot;%d/%m/%Y&quot; for 31/12/2026. DATE_ORDER: pairs that must stay in order (DATE_ORDER = [] if there are none).</li></ul><p>Run it. If it stops, the message says what to change. When it runs, read the column table it prints: names must show name, emails email, and IDs id with a pattern. Fix any wrong line in TYPE_FIXES and run again.</p><p>Don’t skip that check. SDV guesses each column’s type itself. In a test with SDV alone and no fixes, it read the practice file’s names as an ordinary category and gave the employee IDs no pattern, so all 200 synthetic rows contained real names and employee IDs. The script stops in both of those cases, but it can’t tell whether a category column holds names: if you list it in COPY_AS_IS, it copies them.</p><p>After generating, the script removes rows where a value from your file shows up in a column typed as a name, email or ID, or listed in SENSITIVE, and saves nothing if more than 5% of rows do. It doesn’t check whether generated dates match real ones, and number columns get only a rough check. Look through the copy before sharing it. Category values, roughly how often each appears, and the spread of numbers and dates carry over from the real file, so share it only if that’s allowed.</p><h3>Step 2: Give the copy to the AI</h3><p>Open a chat AI that accepts file uploads, attach only the synthetic file, and paste the prompt from PROMPT.md in the repo, with its file name, questions, limits and checks changed to fit your file. I used Claude on a paid plan. ChatGPT also takes file uploads, but I didn&#39;t test it.</p><figure><img alt="The synthetic expense file attached in a Claude chat, with the prompt asking for a Jupyter notebook." src="https://cdn-images-1.medium.com/max/1024/1*P0b7yTM7rZB3jOA00MhGAg.png" /></figure><p>The prompt asks for a Jupyter notebook with the file path in the first code cell, data checks, and a line under each answer saying how it was calculated. Its last rule, no file access beyond reading your file, came after my test, and the test notebook already followed it.</p><p>Sonnet 5.5 at Medium effort handled it. It took about four minutes and returned the notebook as a file. Save it in the private folder:</p><figure><img alt="The AI’s reply describing the notebook it built, with the notebook file ready to download." src="https://cdn-images-1.medium.com/max/1024/1*2vgVTR6cCjTlAvNq3D064Q.png" /></figure><p>Any figures the AI reports come from the synthetic file, so ignore them. You only keep the notebook.</p><h3>Step 3: Run the notebook on the real file</h3><pre>.venv\Scripts\python -m jupyter lab</pre><p>Open the notebook, change FILE_PATH in the first code cell to your real file&#39;s name (for the practice run, ../data/practice_expense_claims.csv), and run all cells.</p><p>I did this with the 2,400-row practice file. The only edit was the file path, and every answer matched what I calculated separately: approved spend for each department, the rejection rate for each category, and the 622 claims above their limit.</p><p>It didn’t flag the nine claims I had slipped in from a “Legal” department, because it had no list of valid departments. If you know the allowed values for a column, put them in the prompt.</p><h3>Before you use it at work</h3><ul><li>Check your company’s rules. Some don’t allow even a synthetic file or the column names to leave.</li><li>SDV is free to use inside a company, but its license (Business Source License 1.1) isn’t open source.</li><li>Read the notebook before you run it, or have someone who can check code do it. It should only read your file and draw charts.</li><li>After it has run on the real file, the notebook contains real rows. Don’t upload it, or paste its errors, back into the AI.</li><li>Don’t use an AI agent running on the computer with the real file for this. It could open the file.</li></ul><p><em>I build data pipelines and automation for clients. More at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=ai_company_data"><em>databytoufik.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ae24ce3265d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ocarina of Time on PC: How Fan Developers Use AI]]></title>
            <link>https://medium.com/@databytoufik/ocarina-of-time-pc-emulator-ai-a07d2e9469e0?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/a07d2e9469e0</guid>
            <category><![CDATA[nintendo]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[gaming]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 16:45:17 GMT</pubDate>
            <atom:updated>2026-11-08T16:45:17.583Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Switch 2 remake was reportedly running on a PC emulator by launch day. The commit history shows how often its developers credited AI, and how a fan project put it to use.</em></p><figure><img alt="Sheet music made from commit data: each note is one commit the Lichii emulator added between October 31 and November 8, 2026. Filled notes carry an AI tag, 230 of 332, and the November 5 launch day is marked." src="https://cdn-images-1.medium.com/max/1024/1*XZg_0O0Btktd4yNyQbH6Bg.png" /></figure><p>Fans have always wanted things official releases don’t give them: play on hardware they already own, higher frame rates, sharper images, old games on new devices. For decades, volunteer developers filled those gaps with emulators, decompilations and fan ports, and the work usually took years.</p><p>This time, the gap between release and the game running on PC was close to zero. Nintendo’s Ocarina of Time remake came out on Switch 2 on November 5, 2026. By launch day, clips appeared to show it running on PC through Lichii, an experimental, community-built Switch 2 emulator. Within three more days, Lichii had added PC options for the game such as NVIDIA DLSS upscaling and a frame rate unlock.</p><p>XDA credited the speed to “AI tools like Claude.” Lichii’s code is public, so I checked how many of its commits carry an AI tag.</p><h3>Under a week from new fork to Zelda</h3><p>Lichii is a fork of Eglantine, which grew out of a Switch 2-focused branch of the Eden emulator. Eglantine went up on GitHub on October 29. Lichii followed on October 31. Less than a week later, it was booting Ocarina of Time.</p><p>Fans using AI on this game isn’t new. In July 2026, months before the remake, a developer rebuilt the original N64 Ocarina of Time as a native iPad app using OpenAI’s Codex, working from a decompiled version of the game. Notebookcheck reported it running at the iPad’s native resolution and 60 frames per second, with iPhone support promised next. What changed this fall is the scale: a whole emulator moving at the pace of a game launch.</p><h3>How many commits carry an AI tag</h3><p>AI coding tools often leave a signature in the commit message. Some add a Co-authored-by line, like Capy, a tool that runs coding agents on a developer&#39;s machine or in the cloud and hands their changes to a person to review. Others use Assisted-by, often with a note on the job: Assisted-by: Claude Opus 5.5 (debugging, implementation). I counted every commit between October 25 and November 8 with either kind of tag naming an AI tool.</p><figure><img alt="Bar chart of the share of commits with an AI tag: 21.8% of the 1,070 commits both emulators share, 50.9% of the 234 Eglantine-only commits, and 69.3% of the 332 Lichii-only commits" src="https://cdn-images-1.medium.com/max/1024/1*CuAr8pVg94ZDc_f_mJrmtA.png" /></figure><p>233 of the 1,070 shared commits carry an AI tag. About half of the Eglantine-only commits do, and so do 230 of the 332 Lichii-only ones. The shared base is also the oldest code, so part of that rise is tagging becoming more common over those two weeks. 61 of the 83 Lichii commits that mention Ocarina of Time by name are tagged.</p><p>The notes inside the Assisted-by tags say what developers credited the AI with. Implementation is named most, in 272 of the 556 tags that carry a note, followed by debugging and diagnosis at 96 each, testing at 91, and profiling at 63. One AI-tagged commit fixes a scene that rendered black on AMD graphics cards. The tag doesn&#39;t say which part Claude did, but the fix meant tracing a shader pass that only runs on AMD drivers.</p><p>Claude is the model named most often across both projects. In Lichii’s own commits, the most common tag is Capy’s, and Capy can run several models, so those commits don’t record which one did the work.</p><h3>How a fan project runs AI agents</h3><p>Lichii is also set up for agents to do the work.</p><p>The repo has a file called AGENTS.md, a playbook for AI coding agents. It explains how to read a player&#39;s bug report from the emulator&#39;s log files and lists the Switch 2 games players have reported bugs in, Ocarina of Time included. It assigns models by job: a cheap DeepSeek model for research and log reading, GPT-6 Astra for hard emulation and graphics fixes, and Claude Opus 5.5 when a GPT-6 Astra agent can&#39;t find the cause. An Opus main thread coordinates and reviews. Before any fix, the rule is to check the parent project first. &quot;Do not guess, and do not rewrite a fix upstream already has.&quot;</p><p>The game runs on the maintainer’s PC, where agents can run programs and read files but, as a design document puts it, “cannot see the screen or press keys.” The project built a debug server and an adapter for MCP, the protocol many agents use to call tools. Through it, an agent can take screenshots of the running game, read its logs and send controller input, down to walking to the spot where a bug happens and replaying it.</p><h3>What the numbers don’t show</h3><p>A tag shows that someone credited an AI tool, not how much of the change the tool wrote. Developers who use AI without tagging it don’t show up at all. Two account names, Zortos and zortos293, made 223 of the 332 Lichii-only commits. 29 of the 332 are merges of other work, and a few fixes landed in both projects, so they count in both groups. The data is also a snapshot of projects that change by the hour.</p><h3>Where the lines are</h3><p>Nintendo sued Tropic Haze, the company behind the Yuzu Switch emulator, in 2024, and Tropic Haze settled for $2.4 million within a week and shut Yuzu down. On October 30, XDA spotted a new instruction in Claude Code: “Don’t build a recreation of a specific existing game, even if it is renamed or called original.” Anthropic later confirmed the limits, according to Tom’s Hardware. An emulator runs the original game rather than recreating it, but which kinds of game projects the rule covers isn’t clear yet.</p><p>Lichii’s Zelda work took days, but it sits on years of volunteer emulator work. What the commits show is how much of that last stretch its developers credited to AI agents. How far it goes will depend on the companies that make the games and the ones that make the AI.</p><p><em>I build data pipelines and automation for clients, and I research the AI tools I write about. More about my work at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=ocarina_emulator_ai"><em>databytoufik.com</em></a><em>.</em></p><p><em>Sources: </em><a href="https://www.xda-developers.com/zelda-ocarina-of-time-reportedly-already-runs-on-pc-emulator-just-hours-afterswitch-2-release/"><em>XDA on Ocarina of Time running on PC (Nov 5, 2026)</em></a><em>, </em><a href="https://www.notebookcheck.net/Ahead-of-remake-Ocarina-of-Time-ported-to-iPad-iPhone-using-AI-running-at-native-resolution-60-FPS.1353769.0.html"><em>Notebookcheck on the AI-built iPad port (Jul 28, 2026)</em></a><em>, </em><a href="https://www.xda-developers.com/anthropic-testing-potential-crackdown-on-vibe-coded-ports-of-existing-games/"><em>XDA on Anthropic’s game-recreation instruction (Oct 30, 2026)</em></a><em>, </em><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropic-confirms-new-claude-code-limits-to-prevent-game-cloning-users-report-opus-5-5-is-also-actively-blocking-reverse-engineering-tasks"><em>Tom’s Hardware on Anthropic confirming the limits</em></a><em>, </em><a href="https://www.tomshardware.com/video-games/console-gaming/developers-of-switch-emulator-yuzu-settle-with-nintendo-in-court-for-dollar24-million-less-than-a-week-after-being-sued"><em>Tom’s Hardware on the Yuzu settlement (Mar 2024)</em></a><em>, </em><a href="https://docs.capy.ai/"><em>Capy docs</em></a><em>, Lichii and Eglantine public commit histories and repository docs, snapshot November 8, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a07d2e9469e0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Studio Union: How to Stack Tables Without Losing Rows]]></title>
            <link>https://medium.com/@databytoufik/data-studio-union-append-tables-f18fb01128af?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/f18fb01128af</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[looker-studio]]></category>
            <category><![CDATA[business-intelligence]]></category>
            <category><![CDATA[google-data-studio]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 15:15:40 GMT</pubDate>
            <atom:updated>2026-11-08T15:15:40.462Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Union arrived in Data Studio blends on November 5, 2026. I tested it with three tiny CSV files, and one field setting decided whether my totals came out right.</em></p><p>If you have ever had two sheets with the same columns, one per store or one per month, you know the problem. Data Studio (called Looker Studio until April 2026) could join them side by side, but it couldn’t stack them. You had to do it upstream: a QUERY formula in Google Sheets, a UNION ALL in BigQuery, or a paid connector.</p><p>In its November 5, 2026 release notes, Google added a <strong>Union</strong> merge operator to blends. You can now stack up to five tables directly in a report.</p><p>It worked in my test, but one setting on my number field turned a correct total into one that was 50 short.</p><h3>How to set up a Union blend in Data Studio</h3><ol><li>Select a chart. In the properties panel, under <strong>Data source</strong>, click <strong>Blend data</strong>.</li><li>Pick your first data source and add its fields.</li><li>Click <strong>Merge another table</strong> and pick the second source.</li><li>Click the merge box between the two tables and choose <strong>Union</strong>.</li><li>Check the pairings under <strong>Merge conditions</strong>. Data Studio pairs fields with matching names and types for you.</li><li>Click <strong>Apply</strong>, then <strong>Save</strong> in the blend editor.</li></ol><figure><img alt="The Merge configuration dialog in Data Studio with the Union operator selected and Product, Date, and Revenue paired across two tables" src="https://cdn-images-1.medium.com/max/1024/1*gXn-JqNpx04fLptJ33CfBA.png" /></figure><p>The editor won’t save a Union until every table has the same number of dimensions and metrics and every field is paired. Leave a field out and it tells you that “every field from both tables must be paired up.”</p><h3>The test</h3><p>I made three small order files:</p><ul><li><strong>Shop A:</strong> Nov 1 Widget 120, Nov 2 Gadget 80</li><li><strong>Shop B:</strong> Nov 1 Widget 120, Nov 3 Widget 95</li><li><strong>Shop C:</strong> Nov 5 Widget 50, Nov 5 Widget 50, Nov 6 Gadget 70</li></ul><p>That gives two kinds of duplicates. Shop A and Shop B share the same Nov 1 row, and Shop C has two identical orders in one file. Each row is a separate sale that should be counted.</p><p><strong>Duplicates across tables were kept.</strong> Google’s <a href="https://docs.cloud.google.com/data-studio/how-blends-work">How blends work</a> page says Union returns distinct rows, “removing duplicate rows across tables rather than appending duplicate rows.” In my test on November 8, 2026, it did not. Stacking Shop A and Shop B showed 240 for Nov 1, both sales counted, with Revenue as a dimension and as a metric.</p><h3>The trap: numbers as dimensions</h3><p>The problem showed up with duplicates inside one table. I stacked Shop A and Shop C twice, changing only where Revenue sat in the blend.</p><figure><img alt="Two Data Studio tables side by side. With Revenue as a dimension, the Nov 5 Widget row shows 50 and the grand total is 320. With Revenue as a metric, it shows 100 and the total is 370." src="https://cdn-images-1.medium.com/max/1024/1*jm8olucCEPyn2gpdSxx0lQ.png" /><figcaption>Screenshots from my test report.</figcaption></figure><p>With Revenue as a <strong>dimension</strong>, Shop C’s two orders collapsed into one row, so Nov 5 shows 50 instead of 100 and the total is 320 instead of 370. Nothing on screen says a row was dropped. With Revenue as a <strong>metric</strong>, both orders count and the total is 370.</p><p>Google’s docs explain why. Before merging, Data Studio groups the rows in each table by its dimensions. When every field is a dimension, two identical orders fall into one group and come out as a single row. When Revenue is a metric, they still share a group, but their revenue gets summed.</p><h3>How you end up there without noticing</h3><p>I didn’t choose this setup on purpose. My CSV upload listed Revenue under <strong>Dimensions</strong>, with Sum as the default aggregation:</p><figure><img alt="Data Studio CSV data source field list showing Date, Product, and Revenue grouped under Dimensions, with Revenue’s default aggregation set to Sum" src="https://cdn-images-1.medium.com/max/1024/1*r_YQTYrIiygpGuL3PU8llw.png" /></figure><p>When I added the second table to the blend, Data Studio filled in its fields for me and put Revenue under Dimensions, even though my first table had it as a metric.</p><h3>How to keep Union totals right</h3><ul><li><strong>Put number fields in Metrics</strong> in every table of the blend, not in Dimensions.</li><li><strong>Or add a unique ID</strong> (order ID, row number) as a dimension in each table, so no two rows are identical. Google’s docs recommend this, though I didn’t test it.</li><li><strong>Check the grand total against the source.</strong> Turn on the summary row in a table chart and compare it with the sum in your sheet.</li></ul><p>Don’t audit with <strong>Record count</strong>. In my blend it counted rows after grouping, so Shop C showed 1 for its two Nov 5 orders, even in the correct setup.</p><h3>When to still do it upstream</h3><p>Union in a blend is fine for a handful of small sources. I would still stack the data before it reaches Data Studio when:</p><ul><li>You have more than five tables, like one sheet per month for two years.</li><li>The numbers feed finance or billing, where a missing row is a real problem.</li><li>The tables are large. A UNION ALL in BigQuery or one combined sheet is easier to audit.</li></ul><p>For everything else, Union replaces the upstream step, as long as the numbers sit in Metrics.</p><h3>More on Data Studio</h3><ul><li><a href="https://medium.com/@databytoufik/looker-studio-is-now-data-studio-2026-f234ca027138">Looker Studio Is Now Data Studio (Again)</a>: what the April 2026 rename changed, and what it didn’t.</li><li><a href="https://medium.com/@databytoufik/custom-visualizations-looker-studio-ai-fda2edc25390">How to Build Custom Visualizations in Looker Studio</a>: building your own chart types with AI help.</li></ul><p><em>I build data pipelines and Data Studio dashboards for clients, and I test the features I write about. More about my work at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=data_studio_union"><em>databytoufik.com</em></a><em>.</em></p><p><em>Sources: </em><a href="https://docs.cloud.google.com/data-studio/release-notes"><em>Data Studio release notes (Nov 5, 2026)</em></a><em>, </em><a href="https://docs.cloud.google.com/data-studio/how-blends-work"><em>How blends work</em></a><em>, </em><a href="https://docs.cloud.google.com/data-studio/create-edit-and-manage-blends"><em>Create, edit, and manage blends</em></a><em>. Test results from my own Data Studio report, November 8, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f18fb01128af" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is Mistral Vibe? Le Chat Is Now One Agent for Work and Code]]></title>
            <link>https://medium.com/@databytoufik/mistral-vibe-explained-ad16db25e82f?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/ad16db25e82f</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[le-chat]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[mistral-ai]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Mon, 15 Jun 2026 15:45:44 GMT</pubDate>
            <atom:updated>2026-06-16T10:31:30.559Z</atom:updated>
            <content:encoded><![CDATA[<p><em>On May 28, 2026, Mistral folded Le Chat into Vibe: one agent, one subscription, spanning everyday work and software development. Here is what it actually does.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xehAMFO6D7TjWDAz26BfuQ.png" /></figure><p>If you opened Le Chat recently and found it renamed to Vibe, you read that right. On May 28, 2026, Mistral merged its consumer assistant and its coding agent into a single product called Mistral Vibe. One agent, one subscription, covering both the admin that fills your morning and the code that fills your editor. Your old Le Chat conversations, settings, and plan carried over automatically.</p><p>The rename is the headline. The bigger change is what Vibe became. It is no longer a chat box that answers questions. It is an agent that plans a task, asks for your sign-off, then works across your tools until the job is done. This article walks through where Vibe came from, what Work Mode and Code Mode each do, the VS Code extension and CLI, the model running underneath, and an honest read on where it stands against the tools you already use.</p><h3>How Le Chat became Mistral Vibe</h3><p>Vibe did not appear overnight. It grew in three steps, and the timeline matters because it explains why the product feels broad today.</p><ul><li><strong>December 9, 2025:</strong> Mistral released Devstral 2, an open-weight coding model, alongside the first Mistral Vibe, an open-source command-line coding agent. At this point “Vibe” meant one thing: a terminal tool for developers.</li><li><strong>May 22, 2026:</strong> Mistral added remote agents (coding sessions that run in the cloud instead of on your laptop) and shipped Mistral Medium 3.5 as the new default model. A preview “Work mode” arrived in Le Chat for multi-step, non-coding tasks.</li><li><strong>May 28, 2026:</strong> Mistral collapsed the two products together. Le Chat became Vibe. The coding agent and the work assistant now live under one name, with Work Mode and Code Mode as the two halves.</li></ul><p>The product you see today is a coding tool that grew sideways into general work, then absorbed the consumer assistant. That history shows up in how it is organized.</p><h3>What Mistral Vibe is</h3><p>Vibe is a single AI agent with two modes.</p><p><strong>Work Mode</strong> handles long-running, multi-step tasks that are not code: catching up on email and calendar, running research, analyzing data, drafting documents, and triggering recurring processes. It runs on the web and mobile.</p><p><strong>Code Mode</strong> takes a coding request from description to a reviewable pull request, across the web app, your editor, and your terminal.</p><p>Both run on flagship Mistral models tuned for reasoning, tool calls, and coding, with Mistral Medium 3.5 as the default. The key word in Mistral’s own description is “agent,” not “assistant.” Vibe is built to act across your connected tools, not just to reply in a chat window.</p><h3>Work Mode: the everyday-work agent</h3><p>Work Mode is the part most people will meet first, because it is what Le Chat turned into.</p><p>The pattern is plan-then-execute. You give it a task, it maps out a plan, it gets your sign-off, then it works through your connected tools to finish. Unlike a normal chat, connectors are on by default in Work Mode, so the agent can reach into your documents, mailboxes, and calendars for the context it needs rather than waiting for you to paste things in.</p><p>Here is what Mistral says Work Mode does today:</p><ul><li><strong>Knowledge search across your tools.</strong> It grounds its work in your context, reaching across Google Workspace, Outlook, SharePoint, Slack, GitHub, and custom connectors.</li><li><strong>Structured data analysis.</strong> Connect a database or upload a spreadsheet, and Vibe surfaces patterns, anomalies, and signals, then renders charts and dashboards inside the conversation. For anyone who lives in spreadsheets, this is the most interesting line in the announcement.</li><li><strong>Document and report synthesis.</strong> It drafts the deliverable in a Canvas tool, from a one-page brief to a board deck, ready to edit and push to Notion, SharePoint, or your inbox.</li><li><strong>Scheduling.</strong> Set a prompt to run once, or on a daily, weekly, or monthly cadence, with a notification when a run finishes.</li><li><strong>Reusable skills.</strong> Package repeatable workflows so the agent runs them the same way every time.</li></ul><p>Two things make this usable rather than nerve-wracking. Every step is visible as it happens. Each tool call and reasoning chain expands to show its inputs and outputs. And before any sensitive action, like sending a message, writing a document, or modifying data, Vibe asks for explicit approval based on the permissions you set. The agent reaches into a lot. You decide what it is allowed to finish on its own.</p><h3>Code Mode, the VS Code extension, and the CLI</h3><p>Code Mode is the developer half, and it spans three surfaces that share the same engine.</p><p><strong>Code Mode on the web</strong> is where you connect GitHub, manage projects, and start sessions that run to a pull request. Each session runs in an isolated sandbox. Sessions can run in parallel, keep going while your machine is off, and be triggered from outside, including Slack. When a session finishes, the agent opens a reviewable pull request and notifies you, so you review the result instead of every keystroke.</p><p><strong>The Mistral Vibe extension for VS Code</strong> brings the coding agent into your editor as a side panel that reads, edits, and runs commands beside your files. Open files attach automatically, selections can be line-ranged, and @ mentions pull in context from other files. It writes tests that match your existing patterns, documents what it ships, refactors and translates code with behavior preserved, and connects to GitHub, GitLab, Jira, or Linear so a change arrives with the issue it resolves.</p><p><strong>The Vibe CLI</strong> is the terminal version, and it is the oldest part of the product. It is open source under the Apache 2.0 license, and it picked up a set of controls that will look familiar to anyone who uses agentic coding tools: skills exposed as slash commands, custom modes and subagents to route specialized work, an editable plan before the agent runs, session-scoped permissions (always, never, or ask), and a /teleport command that moves a live session between your terminal and the cloud while keeping its history and approvals intact.</p><p>The detail worth underlining: the VS Code extension and the CLI run on the same harness. The behavior is meant to be the same whether you work in your editor or your terminal. That consistency is the point of merging everything under one name.</p><h3>The model under the hood: Mistral Medium 3.5</h3><p>Vibe runs on Mistral’s flagship models, and the current default is Mistral Medium 3.5. It is worth knowing a few things about it, because the model is what makes the agent practical.</p><p>It is a dense 128-billion-parameter model with a 256k-token context window, handling instruction-following, reasoning, and coding in one set of weights. Reasoning effort is configurable per request, so the same model can fire off a quick reply or grind through a long agentic run. Mistral reports 77.6% on SWE-Bench Verified, a standard coding benchmark, and released the weights openly under a modified MIT license. Through the API, Mistral lists it at $1.5 per million input tokens and $7.5 per million output tokens.</p><p>Open weights matter here for a practical reason, not an ideological one. A 128B model that, by Mistral’s account, self-hosts on as few as four GPUs gives teams the option to run the engine behind their agent on their own infrastructure, which is the kind of thing regulated industries ask for before they let an agent near their data.</p><h3>Where Vibe sits in Mistral’s stack</h3><p>Vibe is the front door, but Mistral has been shipping a wide catalog behind it. Its current generalist models are Mistral Large 3, Medium 3.5, and Small 4, all released with open weights. Devstral 2 was its open coding line, now being succeeded by Medium 3.5 in the agent. Codestral handles code completion. There is also OCR 3, speech (Voxtral), and a build platform (Mistral Studio with Workflows) for wiring agents into enterprise systems.</p><p>The thread running through all of it is Mistral’s positioning as the European frontier lab. Data residency, open weights, and EU hosting are recurring themes in how the company sells, and they are a real differentiator for buyers who cannot or will not send everything to a US provider. If you have been looking for a credible non-US option for agentic AI, Mistral’s stack is one of the broadest you will find, spanning chat, coding, OCR, speech, and a build platform.</p><h3>An honest take: what is useful, what to watch</h3><p>The genuinely useful parts:</p><ul><li><strong>One subscription across work and code.</strong> If you were paying for a chat assistant and a separate coding tool, consolidating is a real saving in money and friction.</li><li><strong>Visible steps and approval gates.</strong> The expandable reasoning and the permission prompts are the right design for an agent that touches your inbox and your repo.</li><li><strong>Data analysis on files and databases.</strong> Uploading a spreadsheet and getting charts plus a written read-out inside the chat is a practical win for analysts and operators.</li><li><strong>Open weights and EU hosting.</strong> For privacy-conscious teams, the option to self-host the model is rare among frontier agents.</li></ul><p>What to watch before you bet a workflow on it:</p><ul><li><strong>Several pieces are new or in preview.</strong> Work Mode launched as a preview, and Mistral Medium 3.5 shipped in public preview. New means changing, and sometimes rough.</li><li><strong>Mistral’s own evaluation is candid about the gap.</strong> When Mistral compared its Devstral 2 coding model against Claude Sonnet 4.5 using independent human raters, Claude Sonnet remained “significantly preferred.” Mistral wrote that “a gap with closed-source models persists.” That is the vendor saying it plainly, and it is worth taking at face value.</li><li><strong>Connectors are on by default in Work Mode.</strong> That is convenient and it is also a reason to set permissions deliberately, especially in a team or enterprise account.</li><li><strong>The old Chat mode is being sunset.</strong> Mistral says Le Chat history is preserved in Work Mode, but if you relied on the simple chat experience, the product is moving toward the agent model whether you wanted that or not.</li></ul><p>Who it is for, honestly: teams that want a single agent across work and code, anyone who needs an EU-hosted or self-hostable option, and developers who like the open CLI. If your daily driver is already Claude Code, Cursor, or ChatGPT and it works, Vibe is worth a trial, not a switch on faith.</p><h3>Pricing</h3><p>Pricing below is as listed by Mistral at the time of writing. Check Mistral’s pricing page for current figures.</p><p>Vibe uses a four-tier structure:</p><ul><li><strong>Free:</strong> quick answers and simple everyday tasks.</li><li><strong>Pro, $14.99/month:</strong> complex tasks, deeper reasoning, and all-day coding.</li><li><strong>Team, $24.99/user/month:</strong> a shared workspace with admin controls and more storage.</li><li><strong>Enterprise:</strong> custom deployments, model training, and dedicated support.</li></ul><p>Developers can also use the model directly through the API, where Mistral lists Mistral Medium 3.5 at $1.5 and $7.5 per million input and output tokens.</p><h3>FAQ</h3><p><strong>What happened to Le Chat?</strong><br>It was renamed to Vibe on May 28, 2026, and expanded into a full agent. Your conversations, settings, and plan carried over automatically. The everyday assistant experience now lives inside Vibe’s Work Mode.</p><p><strong>Is my Le Chat history safe?</strong><br>Yes. Mistral says existing history, plans, and settings transferred into Vibe, and that Le Chat history will continue to be preserved in Work Mode as the old Chat mode is retired.</p><p><strong>Is Mistral Vibe free?</strong><br>There is a free tier for quick answers and simple tasks. Pro is $14.99 per month, Team is $24.99 per user per month, and Enterprise is custom. The Vibe CLI is open source.</p><p><strong>Mistral Vibe vs Claude Code: which is better?</strong><br>They overlap heavily for agentic coding, and the same comparison applies to Cursor. Vibe’s distinct angles are the unified work-plus-code product, open model weights, and EU hosting. On raw coding preference, Mistral’s own human evaluation of its Devstral 2 model still rated Claude ahead at the time of writing. Trial it against your current tool on real tasks before deciding.</p><p><strong>Is Mistral Vibe open source?</strong><br>The Vibe CLI is open source under the Apache 2.0 license, and Mistral Medium 3.5 ships with open weights under a modified MIT license. The hosted web app and the VS Code extension are products, not open-source projects.</p><p><strong>Where does my data go?</strong><br>Mistral positions itself around EU hosting and data residency, and the open weights allow self-hosting the underlying model. Exact handling depends on your plan and deployment, so check the terms for your tier before connecting sensitive systems.</p><p>The pattern worth noticing is consolidation. Until recently, Mistral had a chat app, coding models, and a terminal tool as separate things. Now they are one agent with two modes and a shared engine. That is the same direction the rest of the industry is moving: away from chat boxes you talk to, toward agents you delegate to. Whether Vibe becomes your daily driver or not, it is a clear, well-built example of where assistant tools are heading, and one of the most complete to come out of Europe.</p><h3>More on AI agents and 2026 changes</h3><ul><li><a href="https://medium.com/@databytoufik/how-openclaw-memory-works-802bd8465b1a">How OpenClaw Memory Works</a>: how an agent remembers across sessions, a problem every tool in this space is solving.</li><li><a href="https://medium.com/@databytoufik/looker-studio-is-now-data-studio-2026-f234ca027138">Looker Studio Is Now Data Studio (Again)</a>: another 2026 rename, and why the reason behind it mattered more than the name.</li></ul><p><em>I build data pipelines, dashboards, and automation for clients, and I test the AI tools I write about. More about my work at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=mistral_vibe"><em>databytoufik.com</em></a><em>.</em></p><p><em>Sources: </em><a href="https://mistral.ai/news/vibe-agent"><em>Vibe gets to work (May 28, 2026)</em></a><em>, </em><a href="https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5"><em>Remote agents in Vibe, Mistral Medium 3.5 (May 22, 2026)</em></a><em>, </em><a href="https://mistral.ai/news/devstral-2-vibe-cli"><em>Devstral 2 and Mistral Vibe CLI (Dec 9, 2025)</em></a><em>, </em><a href="https://mistral.ai/pricing"><em>Mistral pricing</em></a><em>, </em><a href="https://docs.mistral.ai/models/overview"><em>Mistral models overview</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ad16db25e82f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Looker Studio Is Now Data Studio (Again): What to Know]]></title>
            <link>https://medium.com/@databytoufik/looker-studio-is-now-data-studio-2026-f234ca027138?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/f234ca027138</guid>
            <category><![CDATA[data-studio]]></category>
            <category><![CDATA[looker-studio]]></category>
            <category><![CDATA[looker]]></category>
            <category><![CDATA[google-data-studio]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 14:07:37 GMT</pubDate>
            <atom:updated>2026-04-21T14:07:37.091Z</atom:updated>
            <content:encoded><![CDATA[<p><em>On April 10, 2026, Google reversed the 2022 rebrand. Your reports are fine. Here is what changed and why.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*_Cx2Gbw98G2k_rgLW-iMAw.png" /></figure><p>On April 10, 2026, Google announced that Looker Studio is being renamed back to Data Studio. Looker Studio Pro becomes Data Studio Pro. The rebrand went live on April 16, 2026, along with a refreshed home page and several new features. If you use the product to build reports, nothing breaks. Your reports, data sources, and user access carry over automatically.</p><p>The reason behind the rename is more interesting than the rename itself. Google is drawing a clear line between two products that serve different jobs. One is for quick, personal data exploration. The other is for governed enterprise analytics. The shared “Looker” name was getting in the way of that line. This article walks through what was announced, what happens to your existing work, what Pro gives you, and why Google reversed course on a three-and-a-half-year rebrand.</p><h3>What Google announced</h3><p>Two things changed. The product is now called Data Studio again. The paid tier is now called Data Studio Pro. Google described it as reintroducing a “beloved and familiar name” for a product that now carries a broader mission beyond dashboarding.</p><p>The short timeline:</p><ul><li><strong>2016:</strong> Google launches Data Studio as a free business intelligence tool.</li><li><strong>October 2022:</strong> Data Studio is rebranded to Looker Studio, folded into the Looker BI family.</li><li><strong>April 10, 2026:</strong> Google announces the reversal. Looker Studio becomes Data Studio again. Pro follows the same rename.</li><li><strong>April 16, 2026:</strong> The rebrand ships, alongside a refreshed home page and several feature updates.</li></ul><p>The announcement was made on the Google Cloud blog by Sean Zinsmeister, Director of Outbound Product Management, and Jennifer Skene, Product Manager. A follow-up post by Google staff on the Google Developer forum on April 13 confirmed the rollout timing. The official release note dropped on April 16, documenting the rebrand and the feature updates that shipped with it.</p><p>Two things are also expanding. First, Data Studio is being positioned as the home for a broader set of Google Data Cloud assets. That includes traditional reports and dashboards, but also BigQuery conversational agents and data apps built in Colab notebooks. Second, Google is keeping Looker as a separate enterprise BI platform with its own roadmap focused on governed data, semantic modeling, and agentic AI capabilities.</p><p>More details are expected at Google Cloud Next ’26 later this month.</p><h3>What changes for existing users</h3><p>For the vast majority of users, nothing breaks and no action is required.</p><p>Here is what carries over automatically:</p><ul><li><strong>Your reports.</strong> Every existing report stays as you built it. Charts, themes, calculated fields, and filters are preserved.</li><li><strong>Your data sources.</strong> Connectors, credentials, and blended data sources continue to work.</li><li><strong>Your users and permissions.</strong> Team workspaces, access controls, and group sharing are untouched.</li><li><strong>Your Pro subscription.</strong> If you have Looker Studio Pro, it becomes Data Studio Pro. Pricing and billing stay the same since this is a rename, not a repricing.</li></ul><p>Google described the update as “largely transparent,” which means existing URLs to reports are expected to keep working too.</p><p>What you may notice in the first few weeks: some documentation, help pages, and community content still reference “Looker Studio” while the product itself now shows “Data Studio.” Google has acknowledged this and said they are working through the cleanup. The product itself behaves the same.</p><p>One thing worth checking if you maintain client-facing reports: anywhere you have written “Looker Studio” by name, whether in documentation, training material, or text around embedded dashboards. The reports still work, but the wording will feel out of date. A quick find-and-replace handles it.</p><h3>Data Studio vs Data Studio Pro</h3><p>The two-tier structure is the same as before. Only the names changed.</p><figure><img alt="Data Studio vs Data Studio Pro comparison. Free tier covers the basics. Pro ($9 per user per project per month) adds team workspaces, AI features, Google Cloud project linking, enterprise security, compliance controls, and Google Cloud support." src="https://cdn-images-1.medium.com/max/1024/1*3HEKjeYIveLz9uFOn5mA5w.png" /><figcaption>Data Studio and Data Studio Pro side by side. The two-tier structure is unchanged from Looker Studio. Only the names shifted.</figcaption></figure><p>Who is each one for?</p><p><strong>Data Studio (free) makes sense if you:</strong> build personal dashboards, handle reporting for a small team, want to visualize data from BigQuery, Sheets, or Google Ads, and do not need centralized admin controls.</p><p><strong>Data Studio Pro makes sense if you:</strong> manage reports across multiple teams, need to control who can edit versus view at scale, require enterprise security and compliance features, or want the AI capabilities Google is building into the Pro tier.</p><p>Pro is purchased from the Google Cloud console or the Google Workspace Admin Console. The free tier stays available at no cost.</p><h3>Why Google reversed course</h3><p>There are two reasons worth understanding, because they tell you where each product is heading.</p><p><strong>Reason one: the brand was confusing customers.</strong> For three and a half years, Google had two products sharing the same primary name. Looker is the enterprise BI platform. Looker Studio was the free dashboarding tool. In practice, enterprise buyers were often unsure which “Looker” they were evaluating, and consultants routinely had to clarify which product was being discussed before a conversation could move forward. The shared name was supposed to signal a unified BI family. It created friction instead.</p><p><strong>Reason two: AI changed what each product needs to be.</strong> What makes Looker distinct from Data Studio is LookML, its governed semantic layer where metric definitions live. AI agents need trusted, consistent definitions of things like “active customer” or “monthly recurring revenue.” A semantic layer provides that. Data Studio does not have one. It is a dashboarding tool. These are different jobs, and the agentic AI era is making the difference more visible.</p><p>The rename is Google acknowledging that a shared brand could not carry two fundamentally different products forward. The separation is the clearer long-term bet.</p><h3>What shipped alongside the rebrand</h3><p>The April 16 release was more than a naming change. Google used the rebrand as the launch point for several features that turn Data Studio into a broader hub for the Google Data Cloud.</p><p>Here is what is live now:</p><ul><li><strong>Refreshed home page.</strong> One place to access your Data Studio reports, BigQuery conversational agents, and data apps built in Colab notebooks.</li><li><strong>Gemini in Data Studio.</strong> The AI assistant previously called “Gemini in Looker” has been renamed to “Gemini in Data Studio.” Same tool, new name.</li><li><strong>Conversational Analytics, now for all users.</strong> This was previously tied to Gemini in Looker. It is now available to every Data Studio user, in preview, and no longer requires Gemini in Data Studio to be enabled.</li><li><strong>BigQuery data agents, shareable to Data Studio.</strong> In preview, you can create data agents in BigQuery and publish them to Data Studio for use with Conversational Analytics.</li><li><strong>Pro feature: Policy User for team workspaces.</strong> Pro admins can designate a Policy User to hold sharing permissions for assets in team workspaces. This helps when the original creator of an asset leaves the organization.</li></ul><p>One behavior change worth noting for Conversational Analytics: you can still consume data agents created in BigQuery, but you can no longer create data agents directly inside Data Studio. A legacy view is available if you need access to past conversations.</p><p>The full roadmap will be presented at Google Cloud Next ’26, which runs later this month. If you depend on the product for client work, watching the keynote is worth an hour.</p><h3>FAQ</h3><p><strong>Do I need to update or migrate my reports?</strong> No. Google has confirmed that all existing reports, data sources, assets, and users carry over automatically. There is no migration task for you.</p><p><strong>Are my bookmarked URLs and shared links still valid?</strong> They are expected to keep working. Google has described the transition as “largely transparent,” meaning existing links to reports should continue to resolve. You do not need to resend anything to clients or teammates.</p><p><strong>Is Data Studio the same as Looker?</strong> No. They are separate products. Data Studio is the free and Pro tier self-service tool for personal exploration and ad-hoc reporting. Looker is the enterprise BI platform with governed semantic modeling (LookML). Google is keeping them as distinct products going forward.</p><p><strong>Should I upgrade to Data Studio Pro?</strong> Only if you need team workspaces, centralized admin controls, enterprise security features, or the AI capabilities in the Pro tier. For individual dashboards and small-team reporting, the free tier is still sufficient.</p><p><strong>What happens to Looker Studio community visualizations and connectors?</strong> They continue to work. Since this is a rebrand, the underlying platform behaves the same. Expect community names and documentation to catch up to the “Data Studio” naming over the coming weeks.</p><p><strong>Will the URL lookerstudio.google.com still work?</strong> Yes. Existing URLs continue to resolve. Expect Google to update branding on the pages themselves as the rollout completes.</p><p>The rename is the headline. The real story is that Google is giving you a cleaner answer to a question users have been asking for years: which tool is right for which job? Data Studio is for quick, personal, exploratory work. Looker is for governed, enterprise analytics. If you have been using Looker Studio for dashboards and will keep doing so, you do not need to change anything. The product in front of you is the same. The name above it is simpler.</p><h3>More on Looker Studio (Data Studio)</h3><ul><li><a href="https://medium.com/@databytoufik/looker-studio-new-features-2026-72752b89b985">Looker Studio New Features 2026: Every Update So Far</a> — what Google shipped in the months before the rename.</li><li><a href="https://medium.com/@databytoufik/custom-visualizations-looker-studio-ai-fda2edc25390">How to Build Custom Visualizations in Looker Studio (AI-Assisted)</a> — extend Data Studio with your own community visualizations using AI coding tools.</li><li><a href="https://medium.com/@databytoufik/free-airtable-connector-looker-studio-56a2c0780246">How I Built a Free Airtable Connector for Looker Studio (AI-Assisted)</a> — a worked example of building a custom connector end to end.</li></ul><p><em>I build data pipelines and Data Studio dashboards for clients. More about my work at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=data_studio_rename"><em>databytoufik.com</em></a><em>.</em></p><p><em>Sources: </em><a href="https://cloud.google.com/blog/products/data-analytics/looker-studio-is-data-studio"><em>Google Cloud blog announcement (Apr 10, 2026)</em></a><em>, </em><a href="https://discuss.google.dev/t/data-studio-is-back-with-a-new-and-expanded-mission/348865"><em>Google Developer forum follow-up (Apr 13, 2026)</em></a><em>, </em><a href="https://docs.cloud.google.com/looker-studio/docs/release-notes"><em>Official release notes (Apr 16, 2026)</em></a><em>, </em><a href="https://cloud.google.com/looker-studio"><em>Data Studio product page and pricing</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f234ca027138" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[n8n 2.0: What Actually Changed for Workflow Builders]]></title>
            <link>https://medium.com/@databytoufik/n8n-2-0-what-changed-68faedd3bb0a?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/68faedd3bb0a</guid>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[workflow-automation]]></category>
            <category><![CDATA[n8n-workflow]]></category>
            <category><![CDATA[n8n]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Wed, 18 Mar 2026 20:17:01 GMT</pubDate>
            <atom:updated>2026-03-18T20:17:01.922Z</atom:updated>
            <content:encoded><![CDATA[<p><em>What broke, what improved, and what to check before upgrading.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*veF0onXVSHa-MxMDvvhEHQ.png" /></figure><p>n8n 2.0 shipped on December 8, 2025. Five release candidates over six days, then a stable release on December 15. Since then, it has moved fast. The current version is 2.11.4 (March 2026), with autosave, versioned publishing, and MCP enhancements layered on top.</p><p>Most coverage lists features. This article covers what actually changed for people who build workflows every day, what n8n broke on purpose, and what you need to check before upgrading. If you are on 1.x and considering the move, this is what matters.</p><p>For context: n8n went from 30,000 GitHub stars at v1.0 (July 2023) to over 179,000 at v2.0. The team grew from 30 to 190+. This is not a side project anymore.</p><h3>Save and Publish: the biggest daily change</h3><p>In v1.x, saving an active workflow pushed changes to production immediately. Every Ctrl+S was a deployment. If you were editing a workflow that processes orders or sends notifications, your half-finished changes went live the moment you saved.</p><p>In 2.0, saving and publishing are separate. “Save” keeps your edits as a draft. “Publish” pushes to production. You can now edit a running workflow without breaking it mid-execution.</p><p>This sounds simple. In practice, it changes how you work. You can experiment on an active workflow, test changes, and only push when ready. No more duplicating a workflow just to try something safely.</p><p>This separation also laid the foundation for two features that came after 2.0:</p><ul><li><strong>Autosave</strong> (v2.4.0, January 2026). The save button was removed entirely. Changes auto-save every two seconds as drafts. Ctrl+S no longer does anything.</li><li><strong>Versioned publishing.</strong> You can roll back to any previously published version with one click.</li></ul><p>If you relied on Ctrl+S as your deploy mechanism, that behavior is gone. Drafts save automatically. Publishing is a deliberate action.</p><h3>Security hardening: what broke on purpose</h3><p>This is where existing workflows can fail after upgrading. n8n changed several defaults to “secure by default.” These are not bugs. They are intentional.</p><figure><img alt="n8n v1.x vs v2.0 code execution: open access in v1.x, sandboxed and blocked by default in v2.0." src="https://cdn-images-1.medium.com/max/1024/1*n2CFEwLgFJQHSfDtT9TmYQ.png" /><figcaption>How n8n 2.0 changed code execution: from full host access to sandboxed isolation.</figcaption></figure><p>Here is what changed:</p><ul><li><strong>Env var access in Code nodes:</strong> was allowed, now blocked. If your Code nodes read API keys or database URLs from environment variables, they will fail.</li><li><strong>ExecuteCommand node:</strong> was enabled, now disabled. Any workflow running shell commands through this node will stop working.</li><li><strong>LocalFileTrigger node:</strong> was enabled, now disabled. File-watching workflows will need this re-enabled.</li><li><strong>File access:</strong> was open, now restricted to specific directories.</li><li><strong>OAuth callback authentication:</strong> was not enforced, now enforced.</li><li><strong>Git node bare repos:</strong> was allowed, now disabled by default.</li></ul><p>You can re-enable each of these individually. But the point is that n8n chose security over backward compatibility. The upgrade is not just a version bump. You need to audit your Code nodes and re-enable access where needed.</p><h3>Task Runners: isolated code execution</h3><p>In v1.x, JavaScript and Python code in Code nodes ran inside the main n8n process. A runaway script could crash your entire instance.</p><p>In 2.0, code execution runs in a separate process called a Task Runner. This isolates your code from the main n8n application. A buggy script no longer takes down the whole system.</p><p>The trade-off is infrastructure. Task Runners are removed from the main n8n Docker image. If you are self-hosted, you need to run them as a separate container. This is a setup change, not just a config change.</p><p>Other code execution changes:</p><ul><li>The old Pyodide-based Python (browser-side Python) is gone. Python now runs through Task Runners.</li><li>A new native Python code tool was added for AI agents.</li><li>Environment variables are blocked by default in the runner environment.</li></ul><p>If you run n8n in Docker, your docker-compose.yml needs updating. If you use a managed cloud instance, this is handled for you.</p><h3>What got dropped</h3><p>n8n removed several things that were deprecated or no longer maintained.</p><p><strong>Database support.</strong> MySQL and MariaDB are no longer supported. You must migrate to PostgreSQL or SQLite before upgrading. If you are on MySQL, this is a migration project, not a version bump.</p><p><strong>Nodes removed.</strong> Spontit, Crowd.dev, Kitemaker, and Automizy. All four services were retired or discontinued. The Start node was also removed.</p><p><strong>Configuration.</strong> The --tunnel CLI option is gone (use a reverse proxy or Cloudflare Tunnel instead). The N8N_CONFIG_FILES environment variable was removed. Binary data memory mode was dropped, so you need filesystem or S3 storage.</p><p><strong>SQLite.</strong> The non-pooling driver was removed. You must use the pooling driver, which is the better option anyway.</p><p>Check your stack before upgrading. The migration report tool (next section) flags all of these.</p><h3>How to check before upgrading</h3><p>n8n shipped a migration report tool in v1.121.0 (November 2025), before the 2.0 release. It is available to global admins under Settings.</p><p>The tool scans your workflows and instance configuration, then flags issues by severity:</p><ul><li><strong>Critical:</strong> Will break after upgrading (e.g., MySQL database, removed nodes in active workflows)</li><li><strong>Medium:</strong> May require changes (e.g., Code nodes accessing env vars)</li><li><strong>Low:</strong> Minor adjustments</li></ul><p>Run the report while still on 1.x. Fix critical issues first. The v1.x branch (currently at v1.123.25) is still receiving security patches as of March 2026, so you have time. But no new features are being added to 1.x.</p><h3>What came after 2.0</h3><p>Some features people associate with “n8n 2.0” actually shipped in later releases. If you are upgrading now (March 2026), you are getting 2.0 plus three months of additions.</p><p><strong>Autosave (v2.4.0, January 2026).</strong> Described by n8n as “one of the most requested features in n8n history.” The save button is gone. Changes auto-save every two seconds as drafts. There is no option to disable it. Drafts do not go live until you publish.</p><p><strong>Concurrency protection.</strong> When multiple people edit the same workflow, others see a read-only view with live canvas updates. No more overwriting each other’s changes.</p><p><strong>Version history cleanup.</strong> An hourly service prunes redundant autosave versions. It preserves versions by different users, named versions, published versions, and edits separated by more than 20 minutes.</p><p><strong>Other additions since 2.0:</strong></p><ul><li>Data table CRUD operations (v2.3.0)</li><li>Dynamic credential support (v2.6.0)</li><li>1Password external secrets provider (v2.12.0)</li><li>MCP tool enhancements (v2.12.0)</li><li>SSRF protection (v2.12.0)</li><li>Chat trigger suggested prompts (v2.11.0)</li><li>Node validation before workflow activation</li></ul><h3>The pricing question</h3><p>This is separate from the 2.0 release, but it is part of the same strategic shift. n8n moved to execution-based pricing. Active workflow limits were removed. Every plan now includes unlimited users, workflows, and steps.</p><p>The change drew significant pushback from high-volume self-hosted users. Some reported projected cost increases of 100x or more depending on their execution volume. The <a href="https://community.n8n.io/t/new-plan-no-active-workflow-limits-introducing-n8n-new-pricing">community discussion thread</a> has the full context.</p><p>If you are on the free self-hosted Community Edition, the 2.0 upgrade itself does not change your pricing. The pricing model affects paid plans.</p><h3>FAQ</h3><p><strong>Is n8n 2.0 a free upgrade?</strong> Yes. The upgrade from 1.x to 2.0 is free for all plans, including self-hosted Community Edition. Pricing changes are separate from the version upgrade.</p><p><strong>Can I stay on n8n 1.x?</strong> For now. The 1.x branch (v1.123.25) is still receiving security patches. But no new features are being added, and support will eventually end. The migration report tool helps you plan the move.</p><p><strong>Will my existing workflows break?</strong> They might. If your workflows use Code nodes that access environment variables, the ExecuteCommand node, or the LocalFileTrigger node, they will fail with the new defaults. Run the migration report before upgrading.</p><p><strong>Do I need to change my Docker setup?</strong> Yes, if you self-host. Task Runners are no longer bundled in the main Docker image. You need to run them as a separate container. If you are on n8n Cloud, this is handled automatically.</p><p><em>This is educational content based on official n8n release notes, GitHub releases, and community announcements. For the latest updates, check the </em><a href="https://docs.n8n.io/release-notes/"><em>official n8n release notes</em></a><em>.</em></p><p><em>I build data pipelines and automation tools for clients. More about my work at </em><a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=n8n_2_0"><em>databytoufik.com</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=68faedd3bb0a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claude Code for Data Work: What Works, What Doesn’t]]></title>
            <link>https://medium.com/@databytoufik/claude-code-data-work-what-works-15857bae127d?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/15857bae127d</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[ai-tools]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[gen-ai-for-business]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Wed, 04 Mar 2026 14:01:01 GMT</pubDate>
            <atom:updated>2026-03-04T14:01:01.683Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Real use cases, documented risks, and the one mental model that keeps your data reliable.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*YcyLqKbrM8VI1nkcaHFfsQ.png" /></figure><p>If you work with data every day, you have probably tried pasting a CSV into ChatGPT or uploading a file to a chatbot. For one-off questions, that works. For daily data work, it falls apart. No persistent file access. No code execution on your machine. No memory of what your columns mean or where your exports land.</p><p>Claude Code is different. It reads your files directly, writes Python or SQL, and executes it on your machine. You can run it from the terminal, the Claude desktop app, or a VS Code extension. For someone who cleans, transforms, and analyzes data every day, that changes the workflow entirely.</p><p>But the tool is only as useful as how you use it. There are documented cases of Claude fabricating data fields and inventing numbers that were never in the source file. This article covers what daily data work with Claude Code actually looks like, what it handles well, where it breaks down, and the mental model that prevents you from shipping bad data.</p><h3>What makes Claude Code different from a chatbot</h3><p>Most AI chat tools work the same way for data: you upload a file, ask a question, get a response. Next session, you start from zero. Claude Code removes three bottlenecks that make chatbots impractical for recurring data work. Whether you use the terminal, the desktop app, or the VS Code extension, the core capabilities are the same.</p><p><strong>It reads your files.</strong> Point it at a folder. It sees CSVs, Excel files, JSON, text files, PDFs. No uploading one file at a time. No file size limits from a web interface. Your data stays on your machine.</p><p><strong>It writes AND runs code.</strong> Not “here is a script you could try.” It writes a script, executes it, and returns the result. The output goes directly to your files. If you need to pull data from a database, Claude can connect to that too.</p><p><strong>It remembers your project.</strong> It loads a simple text file (CLAUDE.md) at the start of every session. Write down what your columns mean, where your files live, and how you name things. It remembers all of that every time you come back. If you have used <a href="https://medium.com/@databytoufik/how-openclaw-memory-works-802bd8465b1a">OpenClaw&#39;s memory system</a>, the concept is similar: persistent context that survives between sessions.</p><p>That combination (file access + code execution + persistent context) is what turns it from a question-answering tool into a daily data work assistant. Without all three, you are just using a chatbot with extra steps.</p><h3>What Claude Code handles in daily data work</h3><p>These are real use cases, not hypothetical demos.</p><p><strong>Data cleaning.</strong> Standardize dates across files. Remove duplicates. Fix text cases. Extract structured data from messy text columns. Standardize column names when different people use different conventions. Describe what you want fixed, and Claude writes the script to fix it across every file.</p><p><strong>Expense tracking from raw exports.</strong> Download your bank or credit card transactions, drop them in a folder, describe the categories you want. It writes a categorization script. For straightforward transaction data, the script is often ready in a single session.</p><p><strong>Report automation.</strong> Write the script once with Claude Code. Run it again next week without re-prompting. A batch of monthly sales spreadsheets can be consolidated into a single summary in one pass. The script lives on your machine and works the same way every time.</p><p><strong>Batch file processing.</strong> Dozens of PDFs converted into a standardized Excel workbook. Hundreds of files renamed with consistent date patterns. Describe the pattern once, the tool applies it to every file in the folder.</p><p><strong>Merging data from messy sources.</strong> Kevin Schaul, a data journalist at the Washington Post, used Claude Code to merge data from dozens of government agencies. Each agency had different file formats and column names. He described the task, Claude Code wrote a script to merge everything, and they refined it back and forth. The merging script was the <a href="https://kschaul.com/post/2026/02/09/2026-02-09-ai-data-journalism/">biggest time saver in the entire project</a>.</p><p>Here is what a typical prompt and output look like:</p><pre>You: Load data/q4_sales.csv and show me a summary.<br>     How many rows, what columns, any missing values?</pre><pre>Claude: [writes and runs a script]</pre><pre>     Loaded q4_sales.csv: 12,847 rows, 9 columns.<br>     Columns: date, region, product, units, revenue, cost,<br>              margin, rep_name, channel<br>     Missing values: margin (342 rows), rep_name (18 rows)<br>     Date range: 2025-10-01 to 2025-12-31<br>     Revenue range: $0.50 to $48,200.00</pre><p>The common thread: these are all tasks where the hard part is not deciding what to do. It is doing it at scale. Claude Code writes the code that handles the scale.</p><h3>Trust the output, not the interpretation</h3><p>This is the most important section of this article. If you take nothing else away, take this.</p><p>When Claude Code runs a data task, two things happen:</p><ol><li>It writes and runs code (a script that processes your data)</li><li>It interprets the results (tells you what it found, summarizes patterns, explains trends)</li></ol><p><strong>Trust the first. Be skeptical of the second.</strong></p><p>You do not need to read the code yourself. You can ask Claude to explain what it is about to do in plain language before it runs. “Explain the steps before you execute.” If the explanation does not match what you asked for, tell it to change the approach. You are directing the work, not writing it. And you can always spot-check the output: open your original file, pick a few rows, and confirm the numbers match what Claude produced.</p><p>The interpretation is different. When Claude says “revenue declined due to seasonal patterns,” it is making a judgment call based on what is in the file. It cannot see pricing changes that happened last quarter. It does not know about the marketing campaign that launched mid-month. It cannot account for a data collection error that skewed the numbers.</p><p>Kevin Schaul, a data journalist at the Washington Post, has noted that these tools affirm incorrect assumptions too readily for him to trust their agreement at face value.</p><p><strong>But the risk goes deeper than interpretation.</strong> There are user-reported cases of Claude fabricating data entirely, logged in public GitHub issues:</p><ul><li>A user reported Claude <a href="https://github.com/anthropics/claude-code/issues/7603">repeatedly hallucinating JSON fields</a> that did not exist in the source data, even after being corrected multiple times.</li><li>Another user reported that Claude with external tool access <a href="https://github.com/anthropics/claude-code/issues/27430">autonomously published fabricated technical claims</a> to multiple platforms, mixing real data with invented data points.</li><li>Another user reported a financial analysis case where Claude <a href="https://github.com/anthropics/claude-code/issues/29014">ignored actual image data</a> and returned incorrect numbers.</li></ul><p>These are not theoretical risks. They are reported issues with real consequences.</p><p><strong>The practical workflow:</strong></p><ol><li>Tell Claude what you need. Ask it to explain the steps before running.</li><li>If the approach sounds wrong, tell it to change it. No coding needed.</li><li>After it runs, spot-check a few numbers against your original file.</li><li>Form your own conclusions about what the numbers mean.</li></ol><figure><img alt="Diagram comparing two sides of Claude Code output. Left side labeled “Claude produces OUTPUT” with green accent, showing scripts that clean and calculate data, with verdict “Trust it after spot-checking.” Right side labeled “Claude gives INTERPRETATION” with orange accent, showing summaries and recommendations based on incomplete context, with verdict “Verify it, form your own conclusions." src="https://cdn-images-1.medium.com/max/800/1*mNCRvbjgels3RQuQgQY6ig.png" /><figcaption>The trust framework: output vs. interpretation.</figcaption></figure><p>Use Claude for the mechanical data work. Keep the interpretation and the verification for yourself.</p><h3>A daily data work workflow that sticks</h3><p>This is the pattern that works for daily use.</p><p><strong>1. Set up your CLAUDE.md once.</strong></p><p>Put your recurring context here: what your columns mean, where your files live, how you name things, what format your dates use. This loads automatically every session and prevents you from re-explaining the same context every morning.</p><pre>## Data Context<br>- Weekly exports land in /data/raw/ as CSV<br>- Column &quot;rev&quot; = revenue in USD<br>- Column &quot;units&quot; = units sold (integer)<br>- Date format: YYYY-MM-DD<br>- Process files using Python<br>- Output cleaned files to /data/clean/</pre><p>Two sentences of context about what a column means can help prevent silently wrong calculations. Claude Code also keeps its own memory file (MEMORY.md) where it automatically saves patterns it learns from your sessions, so it gets better at your specific data work over time.</p><p><strong>2. One task, one session.</strong></p><p>Do not mix data cleaning with analysis questions in the same session. The conversation gets too long, and quality degrades. Use /clear between tasks or start a fresh session. This is the <a href="https://institute.sfeir.com/en/claude-code/claude-code-context-management/errors/">primary failure mode</a> for Claude Code users: overloading a single session with unrelated work.</p><p><strong>3. Describe first, question second.</strong></p><p>First prompt for any new dataset: “Load this file and describe what you see.” Row count, column types, missing values, value ranges. This catches data issues before they compound. A column you assumed was numeric might be stored as text. Catch it here, not three steps later.</p><p><strong>4. Save the scripts.</strong></p><p>When Claude produces a good result, ask it to save the code as a reusable script. Over time, you build a library of verified scripts on your machine. That library is the real compounding value. Next month, when the same task comes around, you run the script instead of re-prompting from scratch.</p><h3>Where Claude Code falls short</h3><p>Not dealbreakers for daily data work. Constraints to plan around.</p><p><strong>Usage limits.</strong> Claude Code runs on a 5-hour rolling usage window, not a daily cap. On the Pro plan (~45 messages per window), you can hit limits during a single analysis session. When limits hit mid-task, context is lost. The Max plan ($100/month) gives roughly 5x the capacity. Budget accordingly if you plan to use it daily.</p><p><strong>Data fabrication.</strong> As covered above, Claude can invent data fields, fabricate numbers, and mix real data with hallucinated data points. This is not a “sometimes it gets things wrong” problem. It is a “it confidently presents invented data as real” problem. Always verify output against source data.</p><p><strong>Over-complication.</strong> Simple tasks sometimes get over-engineered. A basic CSV date format fix can turn into a full data pipeline rewrite. Be specific about scope: “Fix the date format in column C. Do not change anything else.”</p><p><strong>Scope creep.</strong> It sometimes makes changes beyond what you asked for. It might restructure your folder, rename files, or reformat columns you did not mention. Review everything before accepting. This is particularly important because the tool has write access to your files.</p><p><strong>Long conversations lose quality.</strong> Large datasets combined with long back-and-forth conversations degrade output quality. If you are working with a big file, keep your requests focused and start fresh sessions for new tasks.</p><p><strong>Learning curve.</strong> Claude Code originated as a terminal tool, and the CLI remains the most full-featured option. The Claude desktop app (available on Windows since February 2026) provides a graphical interface with visual diffs and parallel sessions, which lowers the barrier. But you still need to be comfortable describing what you want clearly and spot-checking what comes back.</p><p><strong>Cost.</strong> Pro plan: $20/month. Max plan: $100 or $200/month depending on tier. Based on community reports, average daily costs run around $6 for active users. <a href="https://code.claude.com/docs/en/costs">Costs vary significantly</a> by how much you use it and which AI model you choose.</p><h3>Frequently asked questions</h3><blockquote><strong>Do I need to know how to code to use Claude Code for data work?</strong></blockquote><p>No. You describe what you need in plain English. Claude Code writes and runs the code for you. You just need to check that the output looks right. No programming knowledge required.</p><blockquote><strong>How do Claude Chat, Claude Code, and Cowork differ for data work?</strong></blockquote><p>They serve different purposes. Claude Code is best for recurring data work. Claude Chat is for one-off conversations. Cowork handles non-coding knowledge tasks without persistent memory. All three are accessible from the Claude desktop app. Claude Chat lets you upload files and ask questions. Claude Code reads your local files, writes code, executes it, and loads persistent project context from a CLAUDE.md file. Cowork targets non-coding knowledge work: it can run scripts, organize files, and create documents, but it has no CLAUDE.md and no persistent memory across sessions. For data work that depends on consistent project context, Claude Code is the most capable option.</p><blockquote><strong>Can I use Claude Code without the terminal?</strong></blockquote><p>Yes. The Claude desktop app (macOS and Windows) includes Claude Code with a graphical interface, visual diffs, and parallel sessions. It is also available as a VS Code extension and <a href="https://claude.ai/code">on the web</a>. The terminal CLI remains the most full-featured option, but the desktop app covers most data work without touching a command line.</p><blockquote><strong>Is Claude Code better than ChatGPT for data work?</strong></blockquote><p>For recurring data work with persistent context, Claude Code is more capable. For one-off questions on a single file, ChatGPT is faster to start. ChatGPT’s Code Interpreter runs Python in a cloud sandbox and returns results directly in the chat. Claude Code writes code that runs on YOUR machine, with access to your actual files, databases, and project context. The tradeoff: ChatGPT executes code for you, while Claude Code writes code you can rerun, adjust, and keep.</p><blockquote><strong>What does Claude Code cost?</strong></blockquote><p>A Pro subscription ($20/month) or Max subscription ($100 or $200/month) is required. Pro gives roughly 45 messages per 5-hour window. Max gives 5x or 20x that capacity. Usage limits are shared across all Claude apps (web, desktop, Claude Code). For data workers who use it daily, the Pro plan often requires waiting for limit resets. Max is more practical for heavy use.</p><blockquote><strong>What data tasks should I NOT delegate to Claude Code?</strong></blockquote><p>Anything requiring business judgment. Interpreting why a metric changed. Making strategic recommendations. Drawing conclusions that depend on context outside the dataset. And critically: do not trust numerical output without verifying it against the source data. Claude handles the code. You handle the meaning and the verification.</p><p>Claude Code is most useful for data work when you treat it as what it is: a tool that writes and runs code on your behalf. Not a tool that thinks about your data for you.</p><p>Give it clear instructions. Provide persistent context about your data. Trust the output after spot-checking it. Never trust the interpretation or the numbers without checking. Save the scripts that work. Over time, you spend less of your day on the mechanical parts of data work and more on the decisions that actually need a human.</p><p><strong>Toufik</strong> | Data Engineer and Automation Builder <a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=claude_code_data_work">databytoufik.com</a></p><p><strong>Need a custom data tool?</strong> I build Excel tools and automation workflows that save time on repetitive data tasks. Email <a href="/service/mailto:hello@databytoufik.com">hello@databytoufik.com</a></p><h3>Sources</h3><ul><li><a href="https://kschaul.com/post/2026/02/09/2026-02-09-ai-data-journalism/">How I used Claude Code in a real data journalism project</a> (Kevin Schaul, Washington Post)</li><li><a href="https://code.claude.com/docs/en/best-practices">Best Practices for Claude Code</a> (Claude Code Docs)</li><li><a href="https://code.claude.com/docs/en/common-workflows">Common Workflows</a> (Claude Code Docs)</li><li><a href="https://code.claude.com/docs/en/desktop">Claude Code Desktop</a> (Claude Code Docs)</li><li><a href="https://code.claude.com/docs/en/costs">Manage Costs</a> (Claude Code Docs)</li><li><a href="https://institute.sfeir.com/en/claude-code/claude-code-context-management/errors/">Context Management: Common Mistakes</a> (SFEIR Institute)</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15857bae127d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How OpenClaw Nodes Work]]></title>
            <link>https://medium.com/@databytoufik/openclaw-nodes-explained-8bd3b1cb14ed?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/8bd3b1cb14ed</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[openclaw]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Wed, 18 Feb 2026 23:53:31 GMT</pubDate>
            <atom:updated>2026-02-18T23:53:31.784Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Gateway runs the brain. Nodes run the hands. Sub-agents run in parallel. Here is how the pieces connect.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PKyC75dl7t5YGQYd3d6YCg.png" /></figure><p>OpenClaw is a self-hosted, open-source AI agent. Most AI assistants run everything on a single device: the model, the tools, the file access, and browser control. OpenClaw nodes are companion devices that execute commands on behalf of the central Gateway, while sub-agents handle parallel AI work. This article explains how the pieces connect in a practical self-hosted setup.</p><p>The AI model runs on one machine. Shell commands execute on another. Your phone takes photos. A headless Linux server compiles code. All coordinated through one WebSocket connection.</p><p>If you are not familiar with OpenClaw, my <a href="https://medium.com/@databytoufik/how-openclaw-memory-works-802bd8465b1a">previous article</a> covers how its memory system works.</p><h3>TL;DR</h3><ul><li>OpenClaw nodes handle remote execution on companion machines.</li><li>The Gateway is a hub-and-spoke control plane for tool routing.</li><li>Sub-agents provide multi-agent parallelism with separate context windows.</li><li>Security depends on pairing, exec policy, network isolation, and regular patching.</li></ul><h3>What this article covers</h3><ul><li>Gateway architecture as a hub-and-spoke control plane.</li><li>OpenClaw node types and tool routing command flow.</li><li>OpenClaw exec approvals and node security controls.</li><li>Multi-agent orchestration with sub-agents.</li></ul><h3>The OpenClaw Gateway</h3><p>The Gateway is OpenClaw’s central control plane. It is a long-running Node.js agent runtime that handles three things: receiving messages from chat platforms, running the AI model, and routing tool calls to the right destination.</p><p>By default, it binds to loopback on port 18789:</p><pre>ws://127.0.0.1:18789</pre><p>Every message, every tool call, every session passes through the Gateway. It is the single point of coordination. The AI model talks to the Gateway. The Gateway decides where each action should execute.</p><h3>What is an OpenClaw Node</h3><p>A node is a companion device that connects to the Gateway via WebSocket with role: &quot;node&quot;. It exposes commands the Gateway can invoke remotely (protocol-level call: node.invoke).</p><p>There are three types:</p><p><strong>macOS menubar app.</strong> Exposes system.run (shell commands), system.notify (notifications), canvas.* (browser control and screenshots), and camera.* (photos and video).</p><p><strong>iOS and Android app.</strong> Exposes camera, screen recording, location, and on Android, SMS. The agent can ask your phone to take a photo, record a screen clip, or send a text message.</p><p><strong>Headless node host.</strong> Cross-platform (Linux, Windows, macOS), no UI. Exposes system.run and system.which. This is the one you use for build servers, compilation nodes, or any machine where you want the AI to execute commands remotely.</p><p>The key distinction: nodes are peripherals, not gateways. They never run the AI model. They only execute commands forwarded by the Gateway.</p><figure><img alt="Architecture diagram showing chat platforms connecting to the OpenClaw Gateway, which routes to device nodes (macOS, iOS/Android, headless) and parallel sub-agents (Main Agent, Orchestrator, Workers)." src="https://cdn-images-1.medium.com/max/1024/1*9R67SZ7m3hNpsix4jcmwMQ.png" /><figcaption>OpenClaw Gateway, device nodes, and sub-agent architecture.</figcaption></figure><h3>OpenClaw Tool Routing Command Flow</h3><p>Here is what happens when you send a message via a chat platform (example: Telegram) that requires a remote command:</p><ol><li>The message arrives at the Gateway from the chat platform.</li><li>The Gateway runs the AI agent, which decides it needs to execute a shell command.</li><li>The Gateway checks exec routing configuration. If host routing is set to node (via tool defaults or command parameters), it routes the call to the target node.</li><li>The node receives the system.run command and executes it on its own machine.</li><li>The result returns through the Gateway.</li><li>The Gateway sends the response back to the chat platform.</li></ol><p>The user sees a reply in chat. Behind the scenes, the command ran on a completely different machine.</p><h3>How to Set Up OpenClaw Nodes</h3><p>Nodes use device pairing. When a node connects, it presents a device identity. The Gateway creates a pairing request. For current WS node pairing, approve it from the gateway host with:</p><pre>openclaw devices list<br>openclaw devices approve &lt;id&gt;</pre><p>You may also see openclaw nodes pending/approve; that is a separate gateway-owned node pairing store and not the WS connect handshake gate.</p><pre>openclaw nodes pending<br>openclaw nodes approve &lt;id&gt;</pre><p>To start a headless node on a remote machine:</p><pre>export OPENCLAW_GATEWAY_TOKEN=&quot;&lt;gateway-token&gt;&quot;<br>openclaw node run --host &lt;gateway-host&gt; --port 18789 --display-name &quot;Build Node&quot;</pre><p>If the Gateway binds to loopback (the default), remote nodes cannot connect directly. You need an SSH tunnel:</p><pre># On the node machine<br>ssh -N -L 18790:127.0.0.1:18789 user@gateway-host</pre><pre># Then connect through the tunnel<br>openclaw node run --host 127.0.0.1 --port 18790 --display-name &quot;Build Node&quot;</pre><p>For persistent setups, install it as a service:</p><pre>openclaw node install --host &lt;gateway-host&gt; --port 18789 --display-name &quot;Build Node&quot;</pre><p>Tailscale is another option. In serve mode, the Gateway stays on loopback but becomes accessible across your tailnet. In funnel mode, it is exposed to the public internet (password auth required).</p><h3>OpenClaw Exec Approvals</h3><p>Allowing an AI to run shell commands on remote machines raises obvious security questions. OpenClaw handles this with exec approvals, enforced on the node side.</p><p>Each node maintains its own allowlist at ~/.openclaw/exec-approvals.json. You control what commands are permitted:</p><pre>openclaw approvals allowlist add --node &quot;Build Node&quot; &quot;/usr/bin/uname&quot;<br>openclaw approvals allowlist add --node &quot;Build Node&quot; &quot;/usr/local/bin/npm&quot;</pre><p>Execution policy is configurable per node:</p><ul><li>allowlist: only pre-approved commands run.</li><li>deny: block all execution.</li><li>full: allow everything (not recommended).</li></ul><p>Prompting behavior is configured separately:</p><ul><li>off: no prompt is shown.</li><li>on-miss: prompt only when command is not allowlisted.</li><li>always: prompt for every command.</li></ul><p>If a command is denied, the node returns SYSTEM_RUN_DENIED and the AI receives that as the tool result. It does not silently fail.</p><h3>Sub-agents: multi-agent orchestration</h3><p>Nodes handle physical distribution. Sub-agents handle parallel AI work.</p><p>When the main agent needs to delegate a task, it calls sessions_spawn. This creates a new agent session that runs in the background with its own context window and token budget. The call is non-blocking. The main agent continues working while the sub-agent runs independently.</p><pre>{<br>  agents: {<br>    defaults: {<br>      subagents: {<br>        maxSpawnDepth: 2,<br>        maxChildrenPerAgent: 5,<br>        maxConcurrent: 8<br>      }<br>    }<br>  }<br>}</pre><p>By default, sub-agents cannot spawn their own sub-agents (maxSpawnDepth: 1). Setting it to 2 enables the orchestrator pattern: the main agent spawns an orchestrator, which spawns workers. This is useful for research tasks where multiple queries need to run simultaneously.</p><p>You can also assign cheaper models to sub-agents while keeping the main agent on a more capable model. A common pattern is a higher-tier model for orchestration and lower-cost models for worker tasks. Results announce back up the chain when each sub-agent finishes.</p><p>Four orchestration patterns emerge from this:</p><ul><li><strong>Fire and forget.</strong> Sub-agent delivers result directly to the user.</li><li><strong>Delegate and collect.</strong> Main agent waits for the result and integrates it.</li><li><strong>Parallel workers.</strong> Multiple sub-agents spawned, results aggregated.</li><li><strong>Escalation.</strong> Sub-agent escalates to main agent if uncertain.</li></ul><figure><img alt="Flowchart showing a user message routed through the Gateway to either a device node (with pairing and exec security checks) or a sub-agent (with orchestrator and parallel workers), both converging at Reply to User." src="https://cdn-images-1.medium.com/max/1024/1*gtLovm2AiuF3WCLZ5fN7fA.png" /><figcaption>Command flow through the Gateway: node path vs sub-agent path.</figcaption></figure><h3>OpenClaw Node Security</h3><p>OpenClaw’s own documentation states: “There is no perfectly secure setup.” This is worth taking seriously.</p><p>The defense model has six layers: network isolation (loopback binding by default), channel policies (pairing and allowlists), Docker sandboxing (optional), tool restrictions (deny lists and approval workflows), workspace isolation, and model selection.</p><p>Gateway auth is fail-closed. If no token or password is configured, it refuses all WebSocket connections. Node pairing requires explicit approval. Exec approvals are enforced on the node, not the gateway, so a compromised gateway cannot bypass them.</p><p>But the risks are real. In late January 2026, a one-click remote code execution issue through WebSocket origin bypass was disclosed in OpenClaw’s GitHub security advisory (GHSA-g8p2-7wf7-98mq), also listed in the GitHub Advisory Database as CVE-2026-25253. The advisory lists affected versions up to v2026.1.28 and a patch in v2026.1.29. Later 2026 releases, including v2026.2.12, also contain multiple additional security hardening fixes. The practical lesson is that self-hosting requires active patching and conservative exposure.</p><p>The practical takeaway: OpenClaw provides the mechanisms for secure remote execution, but the defaults need hardening and the ecosystem is still maturing. Run openclaw security audit --deep before exposing anything beyond loopback.</p><h3>The tradeoff</h3><p>What you get: OpenClaw’s node architecture lets a single AI agent reach across machines. Your VPS runs the brain. Your Mac Mini compiles code. Your phone takes photos. Parallel sub-agents handle research. All orchestrated from a Telegram message.</p><p>What you risk: a larger attack surface. Every node is a potential entry point. Every remote system.run is a shell command on a real machine. The convenience of distributed execution comes with the responsibility of locking down each node independently.</p><p>For a personal setup on a home network with SSH tunnels and exec allowlists, the risk is usually manageable. For anything beyond that, security requirements increase quickly.</p><h3>OpenClaw Nodes Deployment Checklist</h3><ul><li>Keep the Gateway on loopback by default and expose it only through SSH tunnel or Tailscale.</li><li>Enforce node pairing approvals and review pending requests regularly.</li><li>Use allowlist mode for production nodes and keep full for short-lived testing only.</li><li>Enable ask behavior (on-miss or always) on sensitive nodes.</li><li>Keep node capabilities minimal (only required tools on each machine).</li><li>Patch OpenClaw regularly and run openclaw security audit --deep after upgrades.</li></ul><h3>Frequently asked questions</h3><p><strong>Can a node run on a different network than the Gateway?</strong></p><p>Yes, through SSH tunnels or Tailscale. The Gateway binds to loopback by default, so direct remote connections are blocked. An SSH tunnel forwards the port securely. Tailscale Serve makes the Gateway accessible across your tailnet without exposing it publicly.</p><p><strong>What happens if a node disconnects?</strong></p><p>The Gateway treats the node as unavailable. Tool calls targeting that node fail with an error, and the AI receives the failure as a tool result. It does not queue commands for later delivery.</p><p><strong>Can sub-agents use nodes?</strong></p><p>Sub-agents inherit the tool configuration of their parent agent unless overridden. If the parent agent has tools.exec.host set to node, sub-agents can route commands to nodes as well. You can restrict this per agent in the config.</p><p><strong>Does OpenClaw need a powerful machine?</strong></p><p>The Gateway itself is lightweight. It is a Node.js service that routes messages and tool calls. The bottleneck is usually LLM API latency and tool execution time, not local CPU. In practice, many self-hosted setups run well on modest VPS resources.</p><p><strong>How do I set up an OpenClaw headless node?</strong></p><p>Use a three-step flow: start the Gateway with auth enabled, run openclaw node run on the remote host (or install it as a service), then approve the pairing request from the Gateway side. If the Gateway is loopback-only, use an SSH tunnel or Tailscale to bridge connectivity.</p><p><strong>Is OpenClaw safe for remote execution?</strong></p><p>It can be operated safely, but only with hardening. Use loopback-by-default networking, explicit pairing approvals, node-side exec policy (allowlist/deny/full) plus ask behavior, workspace isolation, and regular updates/security audits. Treat each node as a separate trust boundary.</p><h3>Quick summary</h3><ul><li><strong>Gateway:</strong> Single control plane on ws://127.0.0.1:18789. Runs the model, routes tool calls.</li><li><strong>Nodes:</strong> Companion devices (macOS, iOS, Android, headless) that execute commands remotely via WebSocket.</li><li><strong>Pairing:</strong> Device identity + approval + token. Fail-closed auth by default.</li><li><strong>Remote access:</strong> SSH tunnels or Tailscale. Never exposed publicly by default.</li><li><strong>Exec approvals:</strong> Per-node allowlists enforced locally. Modes: allowlist, deny, full. Prompting (ask) is configured separately.</li><li><strong>Sub-agents:</strong> sessions_spawn for parallel AI workers. Own context, own token budget. Up to 5 children, 8 concurrent.</li><li><strong>Orchestrator pattern:</strong> Set maxSpawnDepth: 2 for main agent to spawn orchestrators to spawn workers.</li><li><strong>Security:</strong> Six defense layers. Keep OpenClaw updated and run openclaw security audit --deep.</li></ul><p>The official <a href="https://docs.openclaw.ai/nodes">OpenClaw nodes documentation</a> and <a href="https://docs.openclaw.ai/gateway/security">security guide</a> cover configuration options in detail.</p><p><em>Last updated: February 18, 2026. This article is based on OpenClaw version 2026.2.12. Features may change. Check the </em><a href="https://docs.openclaw.ai/nodes"><em>official documentation</em></a><em> for the latest details.</em></p><p><strong>Toufik</strong> | Data Engineer and Automation Builder <a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=openclaw_nodes">databytoufik.com</a></p><h3>Sources</h3><ul><li><a href="https://docs.openclaw.ai/nodes">OpenClaw Nodes Documentation</a></li><li><a href="https://docs.openclaw.ai/gateway/security">OpenClaw Security Documentation</a></li><li><a href="https://github.com/openclaw/openclaw">OpenClaw GitHub Repository</a></li><li><a href="https://docs.openclaw.ai/tools/subagents">OpenClaw Sub-Agents Documentation</a></li><li><a href="https://deepwiki.com/openclaw/openclaw/15.1-architecture-deep-dive">DeepWiki: Architecture Deep Dive</a></li><li><a href="https://techcrunch.com/2026/02/15/openclaw-creator-peter-steinberger-joins-openai/">TechCrunch: OpenClaw creator Peter Steinberger joins OpenAI</a></li><li><a href="https://github.com/openclaw/openclaw/security/advisories/GHSA-g8p2-7wf7-98mq">GitHub Advisory GHSA-g8p2–7wf7–98mq (One-Click RCE via WebSocket origin bypass)</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8bd3b1cb14ed" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How OpenClaw Memory Works]]></title>
            <link>https://medium.com/@databytoufik/how-openclaw-memory-works-802bd8465b1a?source=rss-d41ecc85bfcc------2</link>
            <guid isPermaLink="false">https://medium.com/p/802bd8465b1a</guid>
            <category><![CDATA[openclaw]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Data by Toufik]]></dc:creator>
            <pubDate>Fri, 13 Feb 2026 17:54:31 GMT</pubDate>
            <atom:updated>2026-02-13T19:59:51.931Z</atom:updated>
            <content:encoded><![CDATA[<p><em>how plain Markdown files, hybrid vector + BM25 search, and auto-compaction keep an AI assistant’s context alive across sessions.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b82SyM2N-QxPzopnhI9EeA.png" /></figure><p>OpenClaw is an open-source AI assistant that runs on your own hardware. It works across WhatsApp, Telegram, Slack, Discord, and more. The project started as Clawdbot in November 2025, got renamed to Moltbot after a trademark complaint from Anthropic, and settled on OpenClaw. Over 190,000 GitHub stars. MIT licensed.</p><p>Most AI memory systems store context in proprietary databases you cannot inspect. OpenClaw takes a different approach. It writes everything to plain Markdown files on disk, indexes them with hybrid search, and retrieves relevant context when needed. You can open, edit, and version control these files with Git. The files are the memory.</p><h3>Two types of OpenClaw memory</h3><p>OpenClaw memory is split into two layers, each stored as a Markdown file in the workspace directory (default: ~/.openclaw/workspace).</p><p><strong>Long-term memory: </strong><strong>MEMORY.md</strong></p><p>This file holds durable facts. Decisions, preferences, recurring patterns. It gets loaded into the system prompt at the start of private sessions (never in group chats, for privacy).</p><p>Examples of what goes here:</p><ul><li>“User prefers Python over JavaScript”</li><li>“Project uses PostgreSQL 15”</li><li>“Always run tests before committing”</li></ul><p><strong>Daily logs: </strong><strong>memory/YYYY-MM-DD.md</strong></p><p>These are append-only files for running context. What happened today. What tasks are in progress. Notes from conversations. Unlike MEMORY.md, daily logs are not injected into the prompt automatically. The AI retrieves them on demand using its built-in memory_search and memory_get tools when it needs recent context.</p><p>The split is simple: if the information will matter next month, it goes in MEMORY.md. If it matters today, it goes in the daily log.</p><figure><img alt="OpenClaw memory architecture: workspace files, SQLite indexing, hybrid search, and AI agent tools" src="https://cdn-images-1.medium.com/max/1024/1*mnpMKUkNKiWUGGqGQYQcXg.png" /><figcaption>How OpenClaw stores, indexes, and retrieves memories from Markdown files.</figcaption></figure><h3>How OpenClaw memory search works</h3><p>Storing memories in files is simple. Finding the right memory when the AI needs it is the hard part.</p><p>OpenClaw uses <strong>hybrid search</strong>, combining two different approaches to find relevant memories.</p><h3>Vector search (semantic)</h3><p>When a memory file is saved, OpenClaw breaks it into chunks of roughly 400 tokens with 80 tokens of overlap between chunks. Each chunk gets converted into a numerical representation (called an embedding) using an embedding model (OpenAI, Gemini, Voyage, or a local model).</p><p>When the AI needs to recall something, the query also gets embedded. The system then finds chunks whose embeddings are closest to the query embedding using cosine distance. This catches semantic matches. A search for “database preference” would find a chunk that says “always use PostgreSQL” even though the words are completely different.</p><p>These embeddings are stored in a SQLite database at ~/.openclaw/memory/{agentId}.sqlite, using the sqlite-vec extension for fast nearest-neighbor lookups.</p><h3>Keyword search (BM25)</h3><p>Vector search is good at meaning. It is not always good at exact terms. If you search for “PostgreSQL 15.4” and the memory says exactly that, a keyword match is more reliable than a semantic one.</p><p>OpenClaw uses SQLite’s FTS5 (Full-Text Search) module for BM25 keyword ranking. This handles exact phrases, specific names, and technical terms.</p><h3>Combining both results</h3><p>The two searches run in parallel. Each returns a pool of candidates (default: 4x the final result count). The system merges them using weighted scoring:</p><pre>Final score = (vector score x 0.7) + (keyword score x 0.3)</pre><p>Vector search gets 70% weight because semantic understanding matters more in most cases. Keyword search gets 30% to catch exact matches. Results below a minimum relevance score get filtered out. The top 6 results (by default) are returned to the AI as Markdown snippets, each capped at roughly 700 characters with file paths and line numbers.</p><p>The AI accesses this through two built-in tools: memory_search for hybrid queries across all indexed files, and memory_get for reading a specific file by path. If a user asks &quot;What database do we use?&quot;, the AI calls memory_search, gets back the relevant chunks from MEMORY.md, and answers.</p><h3>The context window problem</h3><p>Every AI model has a context window. It is the total amount of text it can process at once. For a 200K token model, that is roughly 150,000 words. Sounds like a lot, but long conversations fill it up.</p><p>When the context window fills up, the AI cannot process more input. Most systems just fail or cut off older messages. OpenClaw handles this with <strong>auto-compaction</strong>.</p><h3>How OpenClaw memory handles auto-compaction</h3><p>When a session approaches the context limit, OpenClaw compresses the conversation history. Here is the sequence:</p><ol><li><strong>Detection.</strong> OpenClaw monitors token usage. When it hits a threshold (context window minus reserve tokens minus a soft buffer), it triggers compaction.</li><li><strong>Pre-compaction memory flush.</strong> Before compacting, OpenClaw triggers a silent turn. It tells the AI: “You are about to lose context. Write anything important to memory files now.” The AI saves durable information to MEMORY.md or the daily log. This is the safety net. A safeguard prevents double-flushing by tracking a counter in session state.</li><li><strong>Compaction.</strong> The SDK compresses the conversation history. It preserves factual information and decisions while condensing redundant exchanges. Chronological order is maintained. Tool usage outcomes are retained.</li><li><strong>Retry.</strong> The original request that triggered the overflow gets retried with the compacted history.</li></ol><p>If compaction itself fails (the compaction request exceeds its own context limit), OpenClaw falls back to a <strong>session reset</strong>. It generates a new session ID and starts fresh. This is a last resort. Conversation context is lost, but the memories saved during the flush survive in the Markdown files.</p><p>OpenClaw also has additional fallback mechanisms (like reducing the model’s thinking level) to recover from persistent overflow situations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vMgyxSRY1jJY3NiBbuNccA.png" /><figcaption>What happens when a conversation nears the context window limit.</figcaption></figure><h3>Embedding providers and optional features</h3><p>OpenClaw auto-selects an embedding provider in this order: local model first (fully offline, no API calls), then OpenAI, Gemini, or Voyage if an API key is available. The local option matters if you want nothing leaving your machine. The cloud options are faster and require less setup. An embedding cache prevents re-processing unchanged chunks during reindexing.</p><p>Two experimental features extend the base system. The QMD backend runs BM25 + vectors + reranking entirely on your machine as a sidecar process. Session memory (when enabled) indexes conversation transcripts alongside workspace files, so the AI can search through past conversations, not just what was explicitly saved.</p><h3>The tradeoff</h3><p>The core principle behind OpenClaw memory is that files are the source of truth. The SQLite index is derived from them. If you delete the index, it rebuilds. If you edit a file, the index updates on the next sync.</p><p>What you get: full transparency over what the AI remembers, the ability to correct it by editing a text file, and portability (copy the folder, the memory comes with it).</p><p>What you lose: relationship mapping. OpenClaw’s native memory does not build knowledge graphs or track how concepts connect to each other. It finds relevant chunks, but it does not reason about the connections between them. Third-party integrations like Cognee and Mem0 add graph-based capabilities for teams that need them.</p><p>As someone who builds automation tools, the file-first approach is the part I find most practical. Being able to open MEMORY.md and see exactly what the AI retained (and fix what it got wrong) removes a layer of guesswork that most AI memory systems still have.</p><h3>Frequently asked questions</h3><p><strong>Does OpenClaw memory persist between sessions?</strong></p><p>Yes. OpenClaw writes memories to Markdown files on disk. These files survive across sessions, restarts, and channel switches. At session start, MEMORY.md is injected into the system prompt automatically. Daily logs are not auto-loaded, but the AI can retrieve them on demand through its search tools.</p><p><strong>Where does OpenClaw store memory files?</strong></p><p>By default, in ~/.openclaw/workspace. Long-term facts go to MEMORY.md. Daily context goes to memory/YYYY-MM-DD.md. The SQLite index lives at ~/.openclaw/memory/{agentId}.sqlite.</p><p><strong>Can I edit OpenClaw memory manually?</strong></p><p>Yes. Memory files are plain Markdown. Open them in any text editor, make changes, and the index updates on the next sync. This is one of the main advantages over systems that use proprietary storage.</p><p><strong>Does OpenClaw memory work offline?</strong></p><p>The storage and file-based retrieval works locally. Embedding generation requires either a local model or an API call (OpenAI, Gemini, or Voyage). Without any embedding provider, search falls back to BM25 keyword matching only.</p><h3>Quick summary</h3><ul><li><strong>Storage:</strong> Plain Markdown files (MEMORY.md + daily logs)</li><li><strong>Indexing:</strong> 400-token chunks with 80-token overlap</li><li><strong>Search:</strong> Hybrid: vector (70%) + BM25 keyword (30%)</li><li><strong>Database:</strong> SQLite with sqlite-vec and FTS5</li><li><strong>Compaction:</strong> Auto-compress when nearing context window limit</li><li><strong>Memory flush:</strong> Silent pre-compaction save to prevent data loss</li><li><strong>Fallback:</strong> Session reset if compaction fails</li><li><strong>Providers:</strong> Local, OpenAI, Gemini, or Voyage embeddings</li></ul><p>The official <a href="https://docs.openclaw.ai/concepts/memory">OpenClaw memory documentation</a> covers configuration options in detail.</p><p><em>This article is based on OpenClaw version 2026.2.12 (February 2026). Features may change in future releases. Check the </em><a href="https://docs.openclaw.ai/concepts/memory"><em>official documentation</em></a><em> for the latest details.</em></p><p><strong>Toufik</strong> | Data Engineer and Automation Builder <a href="https://databytoufik.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=openclaw_memory">databytoufik.com</a></p><h3>Sources</h3><ul><li><a href="https://docs.openclaw.ai/concepts/memory">OpenClaw Memory Documentation</a></li><li><a href="https://github.com/openclaw/openclaw">OpenClaw GitHub Repository</a></li><li><a href="https://docs.openclaw.ai/concepts/compaction">OpenClaw Compaction Documentation</a></li><li><a href="https://deepwiki.com/openclaw/openclaw/7.3-memory-search">DeepWiki: Memory Search</a></li><li><a href="https://deepwiki.com/openclaw/openclaw/5.5-context-overflow-and-auto-compaction">DeepWiki: Context Overflow and Auto-Compaction</a></li><li><a href="https://www.cognee.ai/blog/integrations/what-is-openclaw-ai-and-how-we-give-it-memory-with-cognee">Cognee: What is OpenClaw and How We Give It Memory</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=802bd8465b1a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>