Bluprint

About

AI agents write a lot of code now, and it’s easy to end up owning a repository you can’t explain. Bluprint reads a repository and builds a map of it, one level at a time. The top level shows the steps the program takes, each step opens into its parts, and the bottom level is the code itself. Every claim on the map cites the lines it came from, anything the map couldn’t prove is marked as a guess, and the map keeps up with the code as it changes.

I’m building it with two friends. The map sits beside a workspace of terminals, coding agents and a live preview of the app, so you can read how something works and then ask about it or change it in the same window. It installs with one line on macOS or Linux and runs through the Claude Code you already use, so the map stays on your computer.

Features

• A generation pipeline in stages: a static inventory of the repository, a fact graph of every function and call across 13 languages, then one Claude agent per section that writes its part of the map from the graph and gets fixed by a checker until it passes
• A verification step that reads every claim against the lines it cites, and takes out sentences that fail before anyone sees them
• A benchmark that scores how well a map answers developers’ questions about a repository, compared with searching it with grep
• Incremental updates: when the code changes, every cited line range moves with the edit, only the sections the change touches are rewritten, and the map shows which ones changed
• Large repositories are mapped by folding directories, so even Kafka’s 4,000+ files fit
• Runs on your own Claude Code or on AWS Bedrock, with a concurrency cap, backoff on throttling, a disk cache so a re-run is identical, and an audit of how each map was made
• A Rust backend with real terminal tabs running Claude Code, Codex, opencode or pi, each in a pseudo-terminal with output streamed to the UI from a reader thread
• A live preview of the project’s app, served through a small HTTP proxy written in Rust that injects a script into each page and tunnels the dev server’s WebSocket, with saved sign-ins
• An MCP server that lets agents read the map and screenshot, click and type in that preview, or in an Android or iOS simulator drawn inside the app
• Review in the map: comment on sections and diff hunks, and send them to Claude as one turn