Context Hub

Overview

Coding agents hallucinate APIs. They use deprecated endpoints, invent parameters, and mix up SDK versions — because their training data is months or years stale. Context Hub is Andrew Ng's open-source solution: a curated, versioned documentation registry that agents fetch via CLI instead of guessing from memory.

Released March 9, 2026. 5,764 stars in 4 days. 68+ API docs already contributed. MIT licensed, published as @aisuite/chub on npm.

How It Works

The Agent Workflow

Agent needs to call Stripe API
  ↓
chub search "stripe"        → finds stripe/api
  ↓
chub get stripe/api --lang js  → fetches current JS docs
  ↓
Agent reads docs, writes correct code
  ↓
chub annotate stripe/api "Webhook needs raw body"  → saves for next time
  ↓
chub feedback stripe/api up  → signals quality to authors

The agent runs chub as a CLI command. No special integration needed — any agent with shell access works. Context Hub ships with a SKILL.md that teaches agents when and how to use it.

Two Content Types

Context Hub makes a clean distinction between docs and skills:

Dimension	Docs	Skills
Purpose	Reference knowledge ("what to know")	Behavioral instructions ("how to do it")
Size	Large (10K-50K+ tokens)	Small (under 500 lines)
Lifecycle	Ephemeral, fetched per-task	Persistent, installed into agent
Language variants	Yes (Python, JS, TS)	No (typically language-agnostic)
Version variants	Yes (v1, v2, etc.)	No
Entry file	DOC.md	SKILL.md

Both live in the same registry, fetched with the same chub get command. The CLI auto-detects the type.

The Learning Loop

The most interesting part isn't the docs — it's how agents learn from using them:

1. Annotations — Local notes agents attach to docs. "Webhook verification requires raw body — do not parse before verifying." Persists across sessions, appears automatically on future chub get calls. The agent doesn't have to rediscover workarounds.

2. Feedback — Up/down ratings with labels (outdated, inaccurate, incomplete, well-structured, helpful). Flows back to doc authors. The docs improve based on real agent usage signals.

3. Community contributions — Anyone can submit docs via PR. Content is plain markdown with YAML frontmatter. The repo IS the docs — you can inspect exactly what your agent reads.

This creates a virtuous cycle: agents use docs → annotate gaps → authors fix gaps → docs get better → agents get better.

Architecture

Content repo (markdown + frontmatter)
  ↓ chub build → registry.json + content tree
CDN (serves registry + files)
  ↓ CLI fetches from here
~/.chub/ (local cache + annotations)
  ↓ CLI reads from here
Agent (consumes docs via stdout)

Multiple sources are supported — remote CDN for community docs, local folders for private/internal docs. Everything merges at the CLI level. Trust is signaled via the source field: official, maintainer, or community.

CLI Commands

Command	Purpose
chub search [query]	Search docs and skills
chub get [id] --lang py	Fetch docs or skills by ID
chub annotate [id] [note]	Attach a persistent note
chub annotate --list	List all annotations
chub feedback [id] up/down	Rate a doc (sent to maintainers)

Category Fit

Context Hub sits at the intersection of three existing categories on this site:

• Agentic Skills Frameworks — Uses the SKILL.md standard, serves skills alongside docs. But it's not a methodology framework — it's a content registry.
• Agent Self-Improvement — The annotation/feedback loop is a lightweight form of agent learning. But it's not memory infrastructure like Mem0 or Letta — it's doc-scoped.
• AI Coding Assistants — It's built for coding agents, but it's not a coding agent itself — it's infrastructure that makes them better.

Best fit: new sub-category within Agentic Skills Frameworks — "Agent Knowledge Registries." Tools that serve curated, versioned knowledge to agents on demand. Context Hub is the first major entrant, but the pattern (community-maintained, agent-optimized documentation) will likely spawn more.

Strengths and Limitations

Strengths:

• Solves a real, painful problem (API hallucination)
• Andrew Ng's backing gives it instant credibility and adoption
• Agent-agnostic — works with any coding agent that has shell access
• Clean docs/skills separation with thoughtful design decisions
• The annotation loop is genuine agent learning without fine-tuning
• Community-driven content model scales well
• 68+ APIs already covered in 4 days

Limitations:

• Requires agents to know to use chub (needs prompting or SKILL.md installation)
• Only as good as the contributed docs — community quality varies
• No MCP server yet (CLI-only, so agents need shell access)
• Annotations are local-only — no sharing across agents/machines
• No private doc hosting (local folders only, no team/org registry)

Comparisons

Tool	What it does	Relationship to Context Hub
MCP servers	Give agents tools to execute	Complementary — Context Hub gives knowledge, MCP gives capabilities
SKILL.md	Behavioral instructions for agents	Context Hub serves skills in SKILL.md format + adds docs
Agentic Context Engine (ACE)	Agents learn from execution traces	ACE learns from doing; Context Hub provides knowledge before doing
Web search	Agents search the open web	Context Hub replaces noisy web search with curated, agent-optimized docs
RAG	Retrieve from document stores	Context Hub is specialized RAG for API docs with versioning and feedback

Bottom Line

Context Hub is infrastructure, not an agent. It makes every coding agent better by solving the API hallucination problem at the source — giving agents current, correct documentation instead of relying on stale training data.

The bigger play is the community model: if Context Hub becomes the npm of agent documentation, it creates a network effect where more contributions → better docs → more agent usage → more feedback → even better docs. Andrew Ng's credibility accelerates this flywheel.

The annotation/feedback loop is what separates it from "just a docs repo." Agents that use Context Hub get smarter over time — not through fine-tuning, but through accumulated local knowledge and community-driven quality improvement.

Watch for: MCP server integration, team/org private registries, cross-agent annotation sharing, and whether API providers start maintaining their own Context Hub entries as a first-class distribution channel.

---

Research by Ry Walker Research • methodology