What it is
Context Hub (Andrew Ng; MIT-licensed, npm @aisuite/chub) gives coding agents curated, versioned, language-specific documentation — with agent skills on the roadmap — so they stop hallucinating APIs and forgetting what they learn within a session. All content is open markdown with YAML frontmatter in a public repo (the same shape as an OKF bundle), so you can inspect exactly what an agent reads and contribute back by pull request.
How it works
Designed for the agent, not the human, through a chub CLI: chub search, chub get <id> --lang py|js, chub annotate, chub feedback up|down. Incremental fetch returns only the reference files needed, to avoid wasting tokens. It can be wrapped as an agent skill via a SKILL.md.
Self-improving loop
Annotations are local notes an agent attaches to a doc; they persist across sessions and can be re-injected (--with-annotations, off by default) but are treated as untrusted input. Feedback (up/down ratings) flows back to doc authors, so the curated content improves for everyone — not just locally. This directly addresses the prompt injection risk for retrieved content.
Where it sits in the stack
A content/skills registry layer that serves the discovery and retrieval layer with curated, versioned, trust-annotated context. It is conceptually distinct from MCP (agent-to-tool transport) and A2A (agent-to-agent interoperability) — the concrete instance that resolves the "Context Hub" open question flagged in the source paper.
Relates to: MCP, A2A, Discovery and retrieval, Agent skill, Evaluation note.
Terms
Glossary terms used here: Agent skill, Prompt injection.