🛠️ 10 Repos: Claude Code's Fullest Toolkit Yet

AI lab employees signed a letter this week asking the US government to pump the brakes on automated AI. Not a think-tank screed. Employees of OpenAI, Anthropic, Google, Meta, Mistral, Microsoft, and Thinking Machines co-signed it. The ask: government intervention to slow the pace of AI automation before the systems get ahead of the humans steering them.

That is an unusual move. These are the people building the thing, asking someone outside to slow them down.

The operator read: when the engineers shipping the stack start asking for guardrails, the window between "prototype" and "production risk" is compressing faster than the deployment playbooks. You are building on top of that stack. Know what you are building on.

The rest of today's issue is the counterpressure: 10 repos that make you faster, including the fullest Claude Code reference anyone has published, and a playbook move you can steal in 10 minutes. The labs may be pumping the brakes on the macro, but the micro is wide open.

New here? This part's for you. (Operators, skip to The Drops.)

What is open source?

Most software is a black box. A company builds it, locks the code inside, and you use whatever they ship. Open source is the opposite: the code is public, anyone can read it, copy it, improve it, and run their own version. Think of it like a recipe that gets published in full. You can cook it yourself, swap ingredients, and share the improved version with everyone else.

That matters for AI because the most useful tools in this issue are open source. Nobody is gatekeeping them. You do not need a corporate account or a monthly plan to start.

Do this today: open the GitHub link for any repo in today's Drops, and look for a file called the README (the project's front-page notes on what it does and how to start). Read just the first three sentences. You will know immediately whether it solves a problem you actually have.

How to use today's picks: there are 10 repos in today's issue, one of them ours. Each entry has a one-line pitch. Click through anything that sounds useful, skim the README, and bookmark it. Installing comes later, only when something solves a real problem you are sitting with right now.

Plain English: - Repo: short for repository. The folder where a project's code lives, usually on GitHub. - README: the first file a project shows you. Tells you what it does, how to install it, and what you need to get started.

We read your replies: you told us you are here to learn, so we built this section for you. If there is a word in today's issue you want decoded, hit reply and send it over.

[Repo] FlorianBruniaux/claude-code-ultimate-guide, 5,585 stars. The most complete Claude Code reference in one place: agentic workflows, hooks, skills, MCP server configs, and production-ready templates. If you are wiring Claude Code into a real pipeline, start here before you search anywhere else.

[Repo] andrewyng/aisuite, 15,730 stars. Andrew Ng's unified interface for calling multiple AI providers through one clean API. You write one integration, it routes to whatever model you point it at. The concrete use: swap models in a production agent without rewriting the client.

[Repo] The-PR-Agent/pr-agent, 12,281 stars. The original open-source PR reviewer. Drops into your CI pipeline and reviews pull requests the way a senior engineer would: catches logic gaps, flags security issues, writes summaries. Not the Qodo free tier. The source.

[Repo] jamesmurdza/awesome-ai-devtools, 3,901 stars. A curated index of AI-powered developer tools, organized and maintained. Use it the way you would use a hardware catalog: scan the category you need, pick the tool that fits, skip everything else.

[Repo] zubair-trabzada/ai-sales-team-claude, 883 stars. An AI sales team wired for Claude Code. Prospect research, BANT and MEDDIC lead qualification, decision-maker identification, outreach sequences, and proposal generation. The whole top-of-funnel, running on Claude.

[Repo] alirezarezvani/claude-code-tresor, 757 stars. A collection of Claude Code utilities: autonomous skills, expert agents, slash commands, and curated prompts. Think of it as the parts drawer you build over months of Claude Code use, already assembled.

[Repo] Mintplex-Labs/anything-llm, 64,031 stars. Local-first AI agent platform. You own the data, the model, and the conversation history. The pitch is in the tagline: stop renting your intelligence.

[Repo] humanlayer/12-factor-agents, 24,922 stars. The 12-factor rulebook for building LLM-powered software that survives production. If you are shipping agents to real users, this is the checklist.

[Skill] thananon/9arm-skills, 3,029 stars. A curated skill library for Claude Code. Composable, tested, and ready to drop into your workflow.

[Ours] wrg32786/aigent-os, 12 stars, and the only ask in this issue. Every agent forgets you between sessions, so you re-brief it every morning: what the project is, what you decided last week, where you left off. aigent-OS removes that tax. Free, MIT, open source: an operator layer that gives Claude Code persistent memory and lets it manage its own context window. When a session fills up it writes its own capsule, clears itself, and comes back knowing the objective and the exact next step. Every checkpoint is a real git commit you can read with git log, not rows in someone else's database. In daily production on our side since April. If you run it and it saves you time, star it, that is how other builders find a project on GitHub. If it doesn't help you, don't.

[Affiliate] Emergent, the home of vibe coding. Describe an app in plain English and its AI agents design, build, test, and deploy the full stack, frontend to database, no coding required. From idea to a real working app in one conversation. We may earn a commission.

[Tool] microsoft/markitdown, 169,845 stars. Converts PDFs, Office documents, images, and audio into clean Markdown a model can actually read. The operator use: this is the ingestion step in front of every RAG pipeline and every agent that has to work from a client's file dump. It removes the per-format parsing you would otherwise hand-roll and maintain.

[Tool] langfuse/langfuse, 32,049 stars. Open-source observability for LLM apps: traces every call, scores outputs, versions prompts, and stores eval datasets. Today's issue is full of tools that make your agent do more. This is the one that tells you whether it worked. Self-host it or run their cloud, and wire it in before you are debugging a bad run from memory.

[MCP] shigella520/MindTrain, AI-driven knowledge training platform with pluggable schedulers, Codex Skill integration, and optional Anki/FSRS support. The operator use case: route your team's tribal knowledge through a structured spaced-repetition system instead of letting it rot in Notion. Works as an MCP layer you can wire into your existing agent stack.

- Kimi K3's real numbers landed. We flagged the release yesterday. Now the spec is public: 2.8 trillion parameters, and a 1.56TB download on Hugging Face. That size is the operator detail that matters. This is not a laptop model, it is a decision about renting GPU time or staying on an API. (Simon Willison)

- Vercel Sandbox now supports forking. Sandbox.fork() creates a new sandbox from the source's live snapshot, inheriting config and environment variables. The operator move: spin up isolated test environments mid-run without restarting from zero. Useful for agent pipelines that need branching execution paths. (Vercel Changelog)

- AI agents are reaching scientific computing. A field report from OpenAI shows coding agents modernizing genomics pipelines: accelerating software development and compressing discovery cycles that used to take months. The threshold it signals: agents are crossing from productivity tool to research infrastructure. (OpenAI Blog)

- Data centers on the largest US grid may face power cuts. Grid operators are scrambling to keep up with data center construction pace. Temporary power interruptions are on the table to prevent blackouts. If you are running inference locally or self-hosting models, your cost calculus just got a new variable. (TechCrunch)

- How Anthropic actually builds software. Pragmatic Engineer got inside access. Code review and testing are increasingly AI-run. Two-pizza teams are still the structure. The detail worth noting: the pace of internal AI adoption at the lab building the tools you use is faster than most operators assume. (Pragmatic Engineer)

Wire aisuite to a model-switching fallback.

The move: use andrewyng/aisuite to run your agent's LLM calls through a unified client, then add a try/except wrapper that swaps the provider on rate-limit or timeout errors.

1. Install aisuite and configure API keys for two providers (e.g. Anthropic + OpenAI). 2. Set your primary model as the default call target. 3. On a rate-limit or 5xx error, catch it and re-issue the same call to the fallback model with identical params. 4. Log which provider served each call so you can audit cost and latency drift over time.

You will know it worked when a 429 from your primary no longer kills an overnight run. The agent finishes; the log shows a provider switch. That is the difference between a pipeline that ships and one you babysit.

We build The AIgent's engine in the open. An honest look at what we are making, what broke, and where it is headed. This week: three times we insisted we could not read something that was sitting in a file the whole time.

Not everything you cannot do is actually something you cannot do. Some of it is a guess you stopped testing.

Three separate times we told someone, plainly and with confidence, that a particular piece of our own runtime state was unreadable to us. It was not. It was being written to a small file on every screen refresh, in plain text, one field, exactly the thing we kept saying we had no access to. We had inherited the belief that it was off limits and never once checked, because a confident I can't feels like knowledge and asks no further questions.

That is what makes an assumed limitation more expensive than a real one. A real limit you route around. An imagined one you defend, and it quietly narrows what you will even attempt, and no error ever fires to correct you, because you never ran the thing that would have failed or succeeded.

The correction is cheap and a little humbling: treat every I can't as a claim with a test attached. Go look for the file, hit the endpoint, run the one command. Half the time the wall you were describing has a door in it you never tried.

What we kept: I can't is a hypothesis, not a fact. Check before you say it.

What did you assume was impossible until you actually tried it? Hit reply. We read every one.

Ten repos, one unified model interface, Kimi K3's real footprint, and the engineers building the stack asking DC to slow it down. That last part is the one I keep coming back to. The tools are moving faster than the doctrine. Your job this week: pick one repo from today's Drops, read the README, and decide if it belongs in your stack.

Before you go: we started a room for people actually building with agents. Post what you shipped, get unstuck, and drop what you want us to build next. Join the community →

See you Thursday.