Open Source Radar — August 25: Karpathy's playbook, open-sourced
Today's board has a single quiet fingerprint: Andrej Karpathy's recurring complaints about how coding agents misbehave have been turned into an instructions file people actually install, and his "LLM Wiki" idea now underpins a self-organizing note system. Around them sit the rest of the stack that thread depends on — a coding agent that runs in your terminal, and a router that makes free LLM inference worth stacking. Four fresh projects, none repeats from the week's earlier radars.
Andrej Karpathy Skills (config, ~206k stars) — The biggest mover on the daily board is just one instructions file, but it's the one people are installing everywhere. It takes Karpathy's running list of LLM coding pitfalls — making wrong assumptions silently, bloating code with dead abstractions, refactoring files they were never asked to touch — and compresses them into four working principles: think before coding, simplicity first, surgical changes, and goal-driven execution. Drop it into Claude Code, Cursor, or Codex and your agent is supposed to ask clarifying questions before guessing, keep diffs minimal, and chase verifiable success criteria instead of vague "make it work." At over 200k stars in days, it's the clearest signal yet that prompt discipline is the cheapest quality upgrade on the agent stack.
Codex (Rust, ~107k stars) — OpenAI's terminal coding agent hit daily trending hard, and it's the one that ties today's thread together because it's the kind of agent those discipline files are meant to tame. It runs locally on your machine, plans before it edits, and works whether you sign in with a ChatGPT plan or pass it an API key. It's been a fixture of the agent stack for a while, but the open-source, Apache-licensed CLI keeps drawing stars on the strength of being a serious, lightweight coding agent that doesn't lock you into a heavy harness. You'd reach for it when you want an assistant that can actually run code and verify results in your repo without standing up a whole platform around it.
FreeLLMAPI (TypeScript, ~19k stars) — An OpenAI-compatible proxy that stacks the free tiers of roughly 28 LLM providers — about 4 billion tokens a month across 358 free endpoints — behind a single local server. Point any OpenAI-style client at one address and it routes each request to the best available free model, fails over when a provider throttles you, and tracks per-key usage so you stay under every cap. It's honest about its limits: the maintainer says it's for personal experimentation and learning, not production, and you swap in a paid API before shipping anything real. For prototyping, tinkering, or running coding agents under a strict budget, it turns a pile of per-provider toys into something usable.
claude-obsidian (Python, ~11.5k stars) — A self-organizing "second brain" for Obsidian driven by Claude Code, built directly on Karpathy's LLM Wiki pattern. Drop in any source — a PDF, a webpage, meeting notes — and the agent reads it, links it, and files it into a connected knowledge graph of plain Markdown that stays yours, not locked in a plugin cache or a cloud database. It grounds claims in a source ledger, builds linked pages and Maps of Content, and answers later questions from what the vault already knows instead of starting every session from a blank slate. It's a local-first, MIT-licensed alternative to hosted note apps for anyone who wants their notes to keep compounding in value without giving up ownership of the files.
Worth watching this week: Karpathy's maxims have stopped being tweets and started being shipped tools — the agents people actually use are increasingly governed by files that say "don't assume, don't over-engineer, verify."
If your agent could file everything you read into a knowledge graph you own, would you trust it — or keep taking notes by hand? Tell us in the comments.
Sources: Andrej Karpathy Skills (GitHub) · Andrej Karpathy's coding pitfalls post · Codex (GitHub) · FreeLLMAPI (GitHub) · claude-obsidian (GitHub)
