// the find
andrewyng/context-hub
Context Hub is a CLI tool that gives coding agents access to versioned, community-maintained API docs as markdown files. Instead of an agent hallucinating a library's interface from training data, it fetches a curated doc that reflects the current API. It's from Andrew Ng's group and targets the problem of agents confidently calling methods that don't exist or have changed.
The content-as-code model is the right call — docs live in the repo as plain markdown with YAML frontmatter, so you can inspect exactly what your agent is reading and send a PR when it's wrong. The incremental fetch design (`--file` for specific reference docs, `--full` for everything) is sensible token management for agents working on a budget. The annotations system is a practical local memory layer: agents can note workarounds per-library and re-inject them next session, with an explicit 'untrusted input' flag that's the correct security posture. MCP server is included (`chub-mcp`), so it plugs into any MCP-compatible agent without custom integration work.
The content library is already enormous — hundreds of packages listed — but quality control at scale is the hard part this repo hasn't solved yet; a stale or wrong doc in here is worse than no doc, and the up/down feedback loop alone won't catch subtle errors. The self-improving agent loop described in the README is more aspirational than real: local annotations require the agent to remember to pass `--with-annotations` on subsequent runs, and most agent frameworks don't wire this up automatically. Coverage is heavily Python-skewed despite JS support existing; many entries only have a Python variant, so JS/TS teams will hit gaps quickly. There's no freshness signal — a doc for a library that shipped a breaking v2 looks identical to an up-to-date one, and the agent has no way to know it's reading stale guidance.