The Anthropic Stack's Substack
Articles about The Anthropic Stack's Substack
Your Coding Agent Can Now Buy the Domain
Your Coding Agent Can Now Buy the DomainYour coding agent can now spend your money. Vercel added domain buying to its command line tool and wrote the changelog with agents in mind. An agent can search for name ideas with `vercel domains search`, check whether they're free with `vercel domains check`, look up the cost with `vercel domains price` and then start the purchase with `vercel domains buy`.
Reverse Engineer Any App With REA, Plus Four More
Reverse Engineer Any App With REA, Plus Four MoreThe Shiplogmorluto/rea (GitHub) - Spot a feature in someone else's app and REA lets your agent work out how it was built without the source code, by inspecting native binaries, Electron and JavaScript apps, .NET assemblies and websites. Every conclusion comes back with its evidence and its limitations. That goes a long way to explaining its 33,340 stars. Run `npx rea-agents setup` to register it with your agent as an MCP server. Native work uses Ghidra or Hopper if you've got one installed.
Stop Paying Frontier Prices for Every Yes or No
Something you can use today.
Copilot Can Now Click Through Your Desktop Apps
The biggest AI story this week.
Pi and Four More Repos You Can Install Today
What developers shipped this week.
Two Agents, No Copying
Two Agents, No CopyingThe Weekly BuildPicture a small agency or any client-facing shop where the delivery work happens across a mess of terminals. Two agents running and you're the one bouncing between them by hand.
Fifty Three Photos OpenAI Never Meant to Publish
The biggest AI story this week.
Five Tools That Actually Solve Agent Infrastructure
Five Tools That Actually Solve Agent InfrastructureThe ShiplogLibreChat (GitHub) - Nearly 45,000 stars and it keeps climbing, which tells you this thing has settled into being infrastructure rather than a novelty project. It's a self-hosted interface that talks to Claude, OpenAI, DeepSeek and most of the rest, with agents and MCP support built in. If you're juggling three subscriptions and four browser tabs to get different models to do different jobs, this is the fix. Set aside an evening to deploy it properly.
Moli: A Lighter Browser for Agents That Only Need to Glance
Moli: A Lighter Browser for Agents That Only Need to GlanceThe Weekly BuildMost agent workflows that touch a browser are doing something small. Check a price. Confirm a form went through. Pull three numbers off a page and move on. And yet the default setup for that job is Chromium, running full weight, eating a few hundred megabytes per instance whether it needs to or not.
Six Clones in Two Days Tells You Something
Six Clones in Two Days Tells You SomethingA model called Jev turned up this week. Within two days there were six clones of it on GitHub. Jev does one narrow thing: it decides, classifies, routes or scores. It doesn't write essays or hold a conversation. The pitch from its maker, TypeSafe, is that it runs more than 100 times faster and more than 200 times cheaper than a small frontier model doing the same job.
Cline Goes Fully Autonomous, Blender Gets Its Own MCP Server
What developers shipped this week.
Say Nothing Changed and Mean It
Say Nothing Changed and Mean ItThe Weekly BuildThere was a post this week from someone running Claude against a legal document. No context given, just "check this over". Claude flagged twenty problems. Every single one was wrong. The document had received a small, pre-approved update and nothing else needed touching, but the model didn't know that. So it invented a document's worth of risk to justify being asked.
OpenAI Says It Solved a Millennium Problem
OpenAI Says It Solved a Millennium ProblemOpenAI says one of its models cracked a Millennium Prize problem this week. That's the seven-problem list with a million dollars attached to each one, set by the Clay Mathematics Institute in 2000. In twenty-five years only one has ever been solved by a human. The claim, if it holds, would be the first by a machine.