1000s of PRs later — does your architecture still make sense? Or has it become spaghetti? GyroCompass reads your living architecture from source, blocks drift the moment a commit crosses a boundary, and keeps Claude Code, Cursor, and every AI agent oriented. So your system stays coherent no matter how fast you ship.
Numbers measured by our open, reproducible benchmark — see the results & methodology →
One command onboards any existing repo. The rest is automatic.
GyroCompass parses your source with Tree-sitter across Python, TypeScript, Go, Rust, Java and more. It produces a living architecture model — components, data flows, API surface, tech stack — written to version-controlled YAML and a visual diagram.
Write your invariants once in .gyro/.gyrorules.yaml — layering rules, ADRs, security policies. GyroCompass installs a real git pre-commit hook that blocks violations before they ever reach review. No advisory warnings — actual enforcement.
Connect the MCP server to Claude Code, Cursor, or any MCP-compatible agent. It gets 10 tools: architecture context, blast-radius analysis, compliance checks, ship-readiness audit, and attestation — grounded in your real system, not hallucination.
Tree-sitter parses 15+ languages into components, relationships, data entities, and API surface — in seconds, with no manual input. The model stays current with every gyro analyze --save.
The pre-commit hook runs gyro check and refuses commits that violate invariants. It names the file, the line, and the rule. The violation never reaches code review.
get_context, get_impact, check_compliance, run_audit, get_drift_report and more — so agents plan inside your architecture from line one, not after review.
8 scanners — hardcoded secrets, PII in logs, vulnerable dependencies, SQL injection, unguarded routes, missing tests and CI. Emits an agent-ready punch list: paste it into Claude, get it fixed in one loop.
Security-firstEvery PR gets a drift score, a list of what changed (which component, which primitive, which rule was violated), and a suggested fix. CI posts it as a comment. Critical drift blocks the merge.
On every PREvery drift event is logged into a typed knowledge graph — what changed, which PR caused it, which rule it violated, and what fixed it. Query the evolution of your system over time.
Long-termA polished SVG diagram auto-generated from your source — components color-coded by type, arrows showing dependencies, a drift-evolution timeline, built-in PNG/PDF export. Standalone HTML, no server.
ShareableWorks offline with Ollama. Plug in OpenAI, Anthropic, or your company's internal endpoint. Zero data leaves your machine unless you choose otherwise. MIT licensed — self-host everything.
PrivateOptional Memgraph + Qdrant backend upgrades drift detection to actual call-graph blast-radius. A change in db.query() surfaces every caller up to the API layer. No grep — real graph traversal.
When an agent or developer crosses a boundary — a route touching the database directly, a rule violated, a capability broken — GyroCompass catches it at the PR and at the commit. Not a style warning. A real block.
--no-verify override
GyroCompass installs a real git pre-commit hook. It runs the full enforcement gate — rules check, drift analysis, attestation — and exits non-zero when the code violates your invariants. The agent sees the exact violation, fixes it locally, and commits cleanly.
--no-verify"We went from 'what did the agent actually change?' to knowing exactly which components moved, which rules it respected, and what it broke — before it hit review."
"The MCP integration means Claude Code reads our actual architecture before proposing a plan. It self-corrects when it's about to cross a service boundary. We spend way less time in review."
"We ran the audit on a Lovable-built app before launch. Found a hardcoded Stripe key, PII in logs, and two vulnerable deps. Pasted the output into Claude, fixed in 20 minutes."
One command. Your architecture extracted, hooks installed, audit run.