GitHub Chat

You point GitHub Chat at a public repo and ask whatever you want about it. It answers questions about structure, functions, and behavior in natural language, so you don't have to guess where to start reading. This is the core job, and it works on any repo that fits within the size limit.

It ships as a Model Context Protocol server, so assistants like Cursor and Claude can use it as a tool. You add a short block to your config, and the assistant gains the ability to query repositories as part of its normal workflow. That means you keep working in the editor you already use.

For programmatic use, GitHub Chat exposes two endpoints. One verifies and indexes a repository, and the other handles chat requests, including multi-turn conversations through a messages array. This makes it possible to bake repository understanding into your own internal tools or bots.

Responses are split into three parts: a rationale, a formatted answer, and the repository context the AI used. Seeing the sources lets you check the reasoning against real files instead of trusting a black box. Markdown support keeps the answers readable.

Because it analyzes the whole repository rather than one open file, answers reflect the bigger picture. It can connect a function to where it is called, or a config to the feature it enables, which is exactly the kind of context a single-file view misses.

GitHub Chat focuses on understanding codebases, while Copilot focuses on writing files. It compares itself to Copilot on purpose, framing the two as complements: one for exploring existing code, the other for generating new code. The API access is the main thing Copilot doesn't offer.