How peek works
Think of it as a Wikipedia article your AI writes about your codebase — and a research desk it can dispatch every time you ask.
peek splits cleanly in two: an offline step that builds the map, and an online step that answers questions using it. You only pay to build the map once.
- Index — build the atlas
peek clones your repo into a private working copy and summarizes it bottom-up: each file is summarized, those summaries roll up into a folder summary, and the folders roll up into a single root map. Every node cites the level beneath it, so an agent can skim the whole repo and then drill straight to the relevant lines. We index into a fresh copy and only swap it in once it's finished — you never see a half-written tree.
- Retrieve — find the right context
When you ask something, the agents start from the root map and work downward, pulling in only the summaries that matter. They also have live tools — code search and file reads against the indexed copy — for when a summary isn't enough and they need the actual source.
- Reason — a team, not a lone chatbot
A director breaks your question into sub-questions. Researchers chase them down in parallel against the atlas and tools. A synthesizer writes the answer from scratch — no team chatter, just the evidence — and a reviewer can send it back for one more pass if something's thin. The result is an answer that's argued, not guessed.
- Answer — grounded and cited
You get a synthesized reply with the source files cited inline, streamed as it's written. See Asking questions for how to scope and read those answers.