Andrew Ng drops Context Hub to fix coding agents' outdated API hallucinations

Every developer using AI agents has lived through this headache: you ask your coding assistant to integrate a common API, and it writes flawless-looking code packed with endpoints deprecated six months ago. The agent sounds confident, but it literally does not know what it does not know.

Andrew Ng's team at DeepLearning.AI just launched an open-source CLI project called Context Hub to put an end to this exact loop of hallucination.

At its core, Context Hub (invoked via chub) provides a clean bridge between coding agents and verified API docs. Instead of sending an AI to scrape noisy web pages, it gives the agent curated, versioned markdown files tailored specifically for LLM context windows.

Just fetch the docs via CLI

The workflow is intentionally dead simple. When an agent needs accurate endpoints, it executes a command like chub get stripe/api to pull language-specific, versioned documentation straight into its context.

Many developers have been trying to solve this by stuffing massive custom instruction files or CLAUDE.md setups into their repos. Context Hub essentially takes that chaotic context engineering habit and packages it into a standard CLI pattern.

It works out of the box with any coding agent that has terminal access. For Claude Code users, the team even shipped a ready-to-use SKILL.md file that drops straight into your skills folder.

Giving agents long-term memory

Fetching static docs is only half the battle. Context Hub's smarter feature is how it handles the gaps and workarounds that official docs never mention.

When your agent spends twenty minutes wrestling with a weird edge case and finally finds a fix, it can run chub annotate to store that local note. In future sessions, the agent automatically resurfaces those notes, keeping it from falling into the exact same trap next week.

There is also a built-in upvote and downvote loop where agents can flag documentation quality, feeding data back to maintainers over time. The team hopes to eventually let different AI assistants share what they learn across ecosystems.

Why it matters

We think Context Hub nails a massive blind spot in current AI tooling. While frontier models are getting much better at reasoning, they are still fundamentally frozen in time at their training cutoffs, and handing them raw web searches often introduces more noise than signal.

For engineers, this tool eliminates the constant chore of hand-feeding updated syntax into system prompts. It saves hours of manual debugging for fast-moving SDKs and payment gateways where parameters change every quarter.

The real test will be doc maintenance. A curated knowledge base is only as good as the humans and community members updating it; if niche libraries fall behind, agents will still hallucinate. Still, standardizing context retrieval into a lightweight CLI feels like the exact pragmatic fix developers need right now.