BELIEVE

AI coding agents now change code faster than people can read it. Eyrie answers two questions: what did my agent just change, and what will this edit break? Eyrie is a live map of how code moves through a repository. It parses a checkout locally with tree-sitter (16 language groups), builds a code knowledge graph where every edge traces to a parsed fact with a confidence score, and derives a six-lane architecture (UI → Edge → API → Logic → Data → External). It stays in sync with the source: on every save only the affected region is recomputed, the diff animates into the map, and it is recorded on a change tape. Built during this hackathon: Meta's models as a first-class backend for Ask the map. A plain-language question is answered by Muse, which may only call Eyrie's read-only graph tools (architecture_overview, find_components, explain_component, impact, trace, find_path). Muse never reads source files. Any answer step that cites a component the tools did not return is rejected rather than shown, and every answer names the model and lists the tool calls it made. I also added an OpenAI-compatible backend so the Meta Model API can be used directly, and used Muse Code to extend language support. The same graph is served to coding agents (Claude Code, Codex, Gemini CLI, Cursor, GitHub Copilot and Muse Code) over MCP: 12 tools, wired in with one command (eyrie init-agents). An agent can check impact before an edit and see what moved after it. Local by default: a loopback server with a per-run token, no accounts and no telemetry. AI is off until you opt in per project or per question. Free and open source. Project page: https://www.badhansen.com/eyrie GitHub: https://github.com/eyrie-hq/eyrie