Compass is a local-first knowledge graph engine for code, built in Rust, that indexes project artifacts only when a user or AI agent explicitly triggers it. No background scanning, no file watchers.
The repo landed on Hacker News with 7 points and one comment — barely a blip. But the design choices matter for anyone building agent context pipelines.
- Local-first, no server: The graph lives on disk alongside your repo. Nothing leaves the machine.
- On-demand indexing: No autoscan, no file watcher. The graph builds or rebuilds only when a skill runs.
- Claude Code integration: Skills like
/eh:optimize-contextbuild or rebuild the graph./eh:orchestrateand/eh:work-on-planrefresh it automatically after workers finish, using the list of files they touched. - Dual-licensed: MIT or Apache-2.0.
- Inspired by Graphify: Built in Rust, described as “evolving beyond” its predecessor.
The core bet: AI coding agents need structured context — code relationships, artifact links, dependency graphs — not just raw file dumps stuffed into a prompt. Compass builds that graph locally so agents can query it without a cloud round-trip or a background daemon eating CPU.
NOTE
The no-autoscan design is deliberate. Every refresh is the consequence of an explicit action — a user-invoked skill or an agent completing work. Activation doesn’t touch disk until something actually runs.
That puts Compass in a different lane from Sourcegraph or GitHub code search, which index continuously and serve from a server. The bet is that agents work better with a structured graph they can query locally than with raw file retrieval or server-side search results.
Compass is early — 7 HN points, 1 comment, and the repo is still finding its audience. If you’re running Claude Code on large repos and hitting context limits, the skill-based graph approach is worth a look. The repo is here.