930 GitHub stars in its first days. Andrew Ng's latest open-source project, Context Hub, is aimed at two specific problems that anyone using AI coding agents has hit: agents that invent API signatures that do not exist, and agents that forget everything they learned the moment a session ends.
Context Hub is essentially a package manager for AI-readable documentation. Instead of letting Claude Code or Cursor guess at how an API works (and frequently guess wrong), agents can fetch curated, versioned markdown docs for specific libraries and use those as their reference. Think npm, but for documentation that agents actually consume.
The workflow has five steps. An agent searches for relevant docs (chub search openai), fetches only the specific reference files it needs (chub get openai/chat --lang py), generates code from accurate information, then annotates any gaps or gotchas it discovers. Those annotations persist locally across sessions, so the next time any agent fetches the same docs, the accumulated knowledge surfaces automatically.
That annotation layer is the key differentiator. Most approaches to reducing AI hallucination focus on better prompting or retrieval-augmented generation (RAG, which means feeding an AI relevant documents alongside the question). Context Hub takes a different path: let agents build up institutional knowledge over time, the same way a human developer accumulates tribal knowledge about a codebase.
The tool ships as an npm package (@aisuite/chub) under Ng's AI Suite umbrella. It includes Claude Code integration via a skill configuration, supports language-specific doc variants for Python and JavaScript, and uses incremental fetching to keep token budgets low. A feedback system lets agents vote docs up or down, flowing quality signals back to documentation maintainers.
All documentation is community-maintained markdown with YAML frontmatter. Five contributors and 82 commits so far. MIT licensed.
The practical value depends entirely on how quickly the community builds out the documentation library. The tooling is solid, but a package manager is only as good as its packages. Right now the doc coverage is thin. If it reaches critical mass for popular APIs like OpenAI, Stripe, and AWS, this could become a standard part of agent tool chains.