🛠️ 9 Claude Code Repos You Can Run Today

Monday. The GitHub trending list refreshed this morning and the Claude Code community did not disappoint.

I counted nine repos worth your time before breakfast, including one that turns any technical book PDF into a live Claude Code skill, one that gives Claude persistent memory of its own mistakes, and a value-investing research framework built specifically for Claude Code and Codex. That last one has 5,206 stars and launched this week. The community is building faster than the labs are shipping press releases.

Five signals worth naming before we get into the drops. First: Anthropic's Mythos model is now accessible, but only to a select group of US companies and government agencies, after weeks of White House negotiation. That is not a broad release, it is a controlled one. Second: China's Zhipu GLM 5.2 lands within a percentage point of Opus 4.8 on a key agentic benchmark, is open-source, and is free. The cost floor just moved again.

The move today is simple: open the drops, pick two repos, clone them before lunch.

[Repo] xbtlin/ai-berkshire, a value investing research framework built for Claude Code and Codex. Runs four master methodologies (Buffett, Munger, Duan Yongping, Li Lu) in parallel via multi-agent research. 5,206 stars since launch this week. If you build finance tools, this is the fastest foundation I have seen.

[Repo] opendatalab/MinerU, converts PDFs, Office docs, and complex layouts into clean LLM-ready markdown or JSON. 71,459 stars. The kind of repo that quietly becomes the first node in every document pipeline you build.

[Skill] virgiliojr94/book-to-skill, feed it any technical book PDF and it produces a Claude Code skill you can reference and use while you work. 6,924 stars. The gotcha: quality of the output skill tracks quality of the source text, so start with a dense practitioner book, not a business bestseller.

[Skill] blader/napkin, a Claude Code skill that writes Claude's mistakes to a per-repo markdown scratchpad so the agent stops repeating them. 560 stars. This is the pattern I wished existed after my third session of watching Claude make the same wrong call.

[Skill] AgriciDaniel/claude-blog, a blog skill suite with 30 sub-skills, 5 agents, and a 5-gate delivery contract. 1,226 stars. Built for both Google ranking and AI citation, which is the right two-target frame for 2026.

[Skill] FradSer/dotclaude, a full Claude Code development environment with specialist agents for code review, security analysis, and technical leadership roles. 567 stars. Think of it as a pre-wired team inside your repo.

[Skill] levineam/qmd-skill, 692 stars. A Claude Code skill from the awesome-claude-code list. The snippet does not describe the exact capability, so clone and read the README before wiring it into a live workflow.

[Repo] hpcaitech/Open-Sora, open-source video generation, 29,151 stars. If you are building any content pipeline that needs programmatic video, this is where you start, not a SaaS subscription.

[Repo] leejet/stable-diffusion.cpp, runs SD, Flux, Wan, and other diffusion models in pure C/C++. 6,396 stars. No Python environment to babysit, ships as a binary, runs on modest hardware.

[MCP] tsouth89/conduit, one local gateway that routes all your MCP servers to every AI client (Claude, Cursor, VS Code, Codex) from a single config. Collapses all servers into 3 meta-tools the agent searches on demand, cutting context usage by roughly 90%. Set it up once and every client inherits the full tool surface. The non-obvious move: use lazy discovery to keep your context window clean on long sessions.

[MCP] LING71671/open-reverselab, an open-source reverse engineering lab with a 197-article knowledge base, MCP tools, and a CTF/APK/PE automation toolchain. Agent-native by design. If your work touches security, binary analysis, or competitive product teardowns, this is the MCP to wire in first.

- Anthropic's Mythos is live, narrowly. After weeks of White House negotiations, the Trump administration permitted Anthropic to grant access to its most advanced model to a select group of US companies and government agencies. This is not a broad release. If you are building on Claude, this is the capability ceiling you are working toward, and the ceiling just got government-gated. (Wired)

- China's GLM 5.2 lands within a point of Opus 4.8 on agentic benchmarks, and it is free. Zhipu's open-source model is being adopted by Chinese enterprises already restricted from US frontier models. The cost floor for agentic capability just moved. Any operator whose workflow assumption is "frontier = expensive US API" should re-run that math. (CNBC)

- OpenAI previewed GPT-5.6 Sol, restricted to trusted partners. Stronger on coding, science, and cybersecurity, paired with an advanced safety stack. The detail that matters: it shipped the same week as Anthropic's tiered Mythos release. Two labs, both gating their best models behind trust layers on the same day. (OpenAI Blog)

- Notion is killing its email app because most users run AI agents on their inbox instead. They are going all in on agent-run inbox management. This is the first major product company to publicly retire a feature category because agents replaced it in production. Note the implication: the products you compete against are being rebuilt around agent defaults, not human workflows. (Ars Technica)

- Ford rehired retired engineers after AI fell short. The quote from leadership: "Mistakenly we thought that by just introducing artificial intelligence ... that would produce a high-quality product." The lesson operators keep relearning is that domain depth is not in the model. It is in the person who knows what good looks like. (TechCrunch)

This week's technique: subagents. Spawn specialized agents, route work to the cheapest model that fits, and run them in parallel.

Most operators run Claude Code as a single thread. The non-obvious move is splitting complex work across multiple subagents, each with a narrow role, running concurrently.

1. In your CLAUDE.md (or inline in your prompt), define the orchestrator's job: decompose the task, assign each piece to a named subagent with a specific role and scope. 2. Dispatch with something like: claude --print "You are a security reviewer. Audit only the authentication module. Return severity-scored findings." for each specialist role. Run them in parallel from a shell loop or a simple orchestrator script. 3. Have the orchestrator collect the outputs and synthesize. The subagents do not need to know about each other.

You will know it worked when the same job that used to burn 40 minutes in one long context window finishes in 8, and the outputs are sharper because each agent stayed in its lane.

Claude Code docs: sub-agents

The move: pipe MinerU into a Claude Code subagent to auto-index any PDF knowledge base.

The pattern: you have a folder of PDFs (research, SOPs, product docs). You want Claude to answer questions against them without hallucinating page references. Here is how to wire it in under an hour.

1. Clone MinerU and run your PDF folder through it: magic-pdf -p ./docs -o ./output -m auto. Every doc becomes clean markdown. 2. Drop the markdown files into a context/ directory in your repo and add a CLAUDE.md rule: "Always search context/ before answering questions about our product or process." 3. Optionally, dispatch a subagent with the role "knowledge base retriever" and scope it to the context/ directory only.

You will know it worked when Claude cites a specific document and section instead of making something up. The gotcha: MinerU's table extraction is excellent, but complex multi-column layouts occasionally merge columns. Spot-check those manually before letting agents answer from them.

A new weekly section. The honest story of the business we are building, one lesson at a time. We are starting at the beginning, with how we cut our teeth.

This started with video.

The first thing we built was a tool that writes, generates, and uploads videos to our own YouTube channels, start to finish, without us touching them. It works. One of those channels just cleared YouTube's bar for monetization.

But the videos were never the real lesson. Building that tool showed us the actual problem with using AI as an operator instead of a toy: it forgets. Every session starts from zero.

So things got tangled. The AI would lose the thread between sessions, redo work it had already finished, and confidently make the wrong call because it could not remember the last one. The more we handed it, the more tangled it got. A smarter model did not fix it.

What fixed it was memory. We gave the AI a way to carry what it learns from one session into the next, to remember the last decision instead of starting cold every morning. The moment it could remember, the tangles started to come apart.

That was lesson one of many, and we did not figure it out in a vacuum. A lot of how we work is staying close to what other builders are doing: finding the open-source projects and repos where people are solving these same problems, and folding what fits into our own system. That habit is becoming the whole point of The AIgent. A resource for people like us, builders figuring this out in real time, sharing what we find and what we learn.

The gap between an AI demo and an AI that can actually run a business is not intelligence. It is the unglamorous infrastructure underneath, built one lesson at a time. Memory was the first. We are going to show you the rest honestly, the wins and the tangles both.

Hitting the same wall, an AI that forgets everything the moment you close the tab? Hit reply and tell us where it is breaking for you. We read every one.

Nine drops, two MCPs, five signals, one playbook move. The through-line today: the community is shipping faster than the news cycle. The repos that matter are on GitHub, not in a press release.

See you Tuesday.