Rig reads your entire codebase to build its graph and runs a local server that exposes it over MCP and HTTP. We take that trust seriously.

• Everything is local. The graph lives in .rig/rig.db inside your project. Embeddings run on-device. No code, query, or graph data is uploaded anywhere.
• The web map binds to loopback (127.0.0.1) by default and requires a per-project session token (.rig/.runtime) even there, because any web page open in your browser can reach a loopback server. rig start opens the map already signed in; your agent's rig_session_token tool hands you a login link. rig serve --no-auth turns the token off, but only on loopback: Rig refuses to bind a non-loopback address without one.
• The map server also refuses requests whose Host isn't one of its own names (DNS rebinding) and writes or WebSocket connections from any origin but its own (CSRF), and it sends no CORS headers.
• rig install writes MCP server entries into your agent's config files (e.g. .claude.json). It merges, never clobbers.
• The first run downloads an embedding model from the Hugging Face CDN to your local model cache. This is the only outbound network call in the core flow.

Until a 1.0 release, only the latest published version receives security fixes.

Please do not open a public issue for security reports.

Email kingolajohn55@gmail.com with:

• a description of the issue and its impact,
• steps to reproduce (a minimal repro is ideal),
• any suggested remediation.

We aim to acknowledge within 72 hours and to ship a fix or mitigation before public disclosure. We'll credit you in the changelog unless you'd prefer to remain anonymous.

In scope: the rig CLI, the MCP server, the web server/auth, the installer's config writes, and anything that could exfiltrate local code or graph data.

Out of scope: vulnerabilities in upstream dependencies (report those upstream), and issues requiring a pre-compromised local machine.