Two fundamental failures plague every AI coding agent:

What this looks like in practice

# Agent confidently writes:
from stripe import PaymentProcessor

# This class doesn't exist.
# The real import is:
import stripe
stripe.PaymentIntent.create()

# Agent discovers this, fixes it.
# Next session? Same mistake.

• A CLI tool (@aisuite/chub) published to npm
• A content repository of markdown documentation with YAML frontmatter
• A feedback system connecting agent usage to content quality
• A local annotation store for persistent cross-session learning

How agents find documentation

# Broad search
$ chub search "stripe"

Found 3 results:
  stripe/api        Stripe API reference
  stripe/webhooks   Webhook handling guide
  stripe/checkout   Checkout integration

# List everything available
$ chub search

All available content (47 items):
  openai/chat       OpenAI Chat Completions
  openai/embeddings Embedding API
  stripe/api        Stripe API reference
  ...

Search is the entry point

• Text-based query matching against doc IDs and descriptions
• Returns structured list of available content
• Agents use this to decide which docs to fetch
• Empty query = browse all available content

Retrieve only what you need

# Default: main doc only
$ chub get openai/chat --lang py

# Specific reference file
$ chub get openai/chat --file streaming

# Everything (all references)
$ chub get openai/chat --full

Each doc can have multiple reference files — the agent fetches only the section relevant to its current task.

Why this matters

Approximate token counts for typical API documentation.

Same API, different languages

# Python variant
$ chub get openai/chat --lang py

# JavaScript variant
$ chub get openai/chat --lang js

Documentation is maintained per-language, not as a one-size-fits-all reference. Each variant includes:

• Language-idiomatic code examples
• Correct import paths and package names
• Language-specific gotchas and patterns

How agents learn across sessions

# Agent discovers a gotcha
$ chub annotate stripe/webhooks \
  "Must use raw request body, not
   parsed JSON, for signature
   verification. See: express
   raw() middleware."

# Next session: annotation appears
# automatically when fetching docs
$ chub get stripe/webhooks

[Content + local annotations shown]

# Manage annotations
$ chub annotate --list
$ chub annotate stripe/webhooks --clear

Key properties

• Local storage — annotations live on your machine, not the server
• Auto-surfaced — they appear every time you fetch that doc
• Per-document — each doc ID has its own annotation list
• Agent-written — designed for agents to write, not humans

Community-driven quality

# Agent found docs helpful
$ chub feedback openai/chat up

# Agent found docs incomplete
$ chub feedback stripe/api down

Votes are aggregated and surfaced to content authors. This creates a signal for which docs need improvement.

The feedback loop

Markdown + YAML frontmatter

---
id: openai/chat
title: Chat Completions API
version: 2024.1
languages: [py, js]
references:
  - streaming
  - function-calling
  - vision
---

# Chat Completions

Create chat completions using
the OpenAI API...

## Basic Usage
...

Repository structure

content/
  openai/
    chat/
      index.md        # main doc
      py.md           # Python variant
      js.md           # JS variant
      streaming.md    # reference file
      function-calling.md
  stripe/
    api/
      index.md
      py.md
      webhooks.md

SKILL.md pattern

Agents discover Context Hub through a skill file that teaches them when and how to use it:

# For Claude Code:
mkdir -p ~/.claude/skills/get-api-docs
cp SKILL.md ~/.claude/skills/get-api-docs/

# The SKILL.md tells the agent:
# 1. When to use chub (API questions)
# 2. How to search and fetch
# 3. When to annotate findings
# 4. When to provide feedback

Workflow inside an agent session

Technology stack

• Runtime: Node.js ≥ 18
• Package: @aisuite/chub on npm
• Content: GitHub repo (markdown + YAML)
• Local store: File-based annotation persistence
• Language: 100% JavaScript

Installation

npm install -g @aisuite/chub

# Verify
chub --version
chub help

How fetch works

// Simplified flow:

// 1. Parse doc ID + flags
const { id, lang, file } = parseArgs();

// 2. Resolve content path
const path = resolve(
  `content/${id}/${lang || 'index'}.md`
);

// 3. Fetch from GitHub
const content = await fetchContent(path);

// 4. Merge local annotations
const annotations = getAnnotations(id);

// 5. Return combined output
return format(content, annotations);

Context Hub is part of Andrew Ng's aisuite project — a collection of tools for building reliable AI agents.

• aisuite — unified API for multiple LLM providers
• context-hub — curated docs for agents (this talk)

The philosophy: agents need infrastructure, not just prompts. Better tools make better agents.

Without Context Hub

# Agent hallucinates:
from stripe.checkout import Session

session = Session("price_123")
session.redirect_url = "/success"
session.create()

# ImportError: No module named
# 'stripe.checkout'

# Agent tries 3 more wrong versions
# before getting it right...

With Context Hub

# Agent fetches docs first:
# $ chub get stripe/checkout --lang py

import stripe

session = stripe.checkout.Session.create(
  line_items=[{
    "price": "price_123",
    "quantity": 1,
  }],
  mode="payment",
  success_url="https://...",
)
# Works on first try.

Who should contribute?

• API providers — ship docs that agents can actually consume correctly
• Framework authors — ensure agents use your library's idiomatic patterns
• Developers — fix the gaps and gotchas you discover while coding
• Companies — reduce support tickets by giving agents correct answers

How to contribute

# 1. Fork the repo
git clone andrewyng/context-hub

# 2. Add content
mkdir -p content/myapi/
# Write markdown with YAML frontmatter

# 3. Submit PR
git push && gh pr create

# Content guide in repo explains
# frontmatter format and conventions

Current limitations

• Content coverage — still growing; not every API is documented yet
• Staleness risk — docs can fall behind API changes if not maintained
• Node.js dependency — requires Node 18+ installed on the system
• Discovery — agents need the SKILL.md to know Context Hub exists

Open questions

• How to handle version conflicts when APIs change frequently?
• Can annotations be shared across teams while preserving privacy?
• How to scale curation as the content library grows?
• Should agents be able to auto-contribute docs from their discoveries?

Planned expansions