Faultlines ships an MCP server that Cursor, Claude Code, Cline, and Aider call to get precise codebase context — features, owners, hotspots, live errors from Sentry, and the right symbols. Same data your engineering leaders see — never your raw source code.
Your agent stops reading the wrong files and asking you what owns what. It pulls the feature map, finds the relevant symbols, and answers with the same context you’d hand a senior engineer on day one.
Side benefit: 60–80% fewer tokens per query vs a naive grep-and-read
workflow (~15k → ~3k with--symbols).
Your AI agent gets the feature map, bug hotspots, owners, coverage — plusimpact analysis (which features break, missing co-changes, regression risk),symbol-level precision (read only the functions that matter, not whole files), andlive runtime signal from your Sentry and PostHog accounts.
Your MCP server runs where your privacy posture demands. The tool API stays identical — only the data path changes. Pick once at the org level; switch any time as your compliance bar shifts.
The MCP server runs in our cloud. Your agent calls it over HTTPS with an org-scoped token. Lowest setup friction — works with hosted agents like Claude.ai web in addition to local ones.
The MCP server runs as a local process on your dev’s machine. It pulls encrypted scans from us, decrypts with your org key in memory, and serves the agent with rich data we never see in plaintext.
The entire Faultlines stack — engine, dashboard, MCP — deploys as a Docker image inside your infrastructure. Agents call your internal endpoint. Faultlines.com only sees a license-check ping.
Your agent doesn’t just see static feature shapes. It sees which features are failing in production, which are hot in the dashboard, and which had a coverage drop in the last PR — pulled live from your Sentry and PostHog accounts.
Each Sentry stack trace is matched against the scan from the same
commit. The agent calls
get_feature_errors
and gets
“Billing → 47 errors last week, top:
TypeError in handleCheckout” — not a generic error feed.
Page views and custom events tagged by URL get mapped to features via the scan’s route index. The agent can answer “which feature has the highest bounce” or “is anyone using the new Settings page” without any code instrumentation.
After every push, a Faultlines incremental scan runs (~5 seconds, cents in LLM cost). Your branch’s scan updates, your agent’s context updates. Same applies to PR branches — feature/auth-v2 has its own current scan, separate from main.
Any MCP-compatible client. Drop the server into the config, restart the agent, and start asking questions about your codebase that previously required a senior engineer.
The feature map and ownership graph that powers these tools is the same artifact your EM/CTO uses to decide what to refactor next quarter — and it’s the same data the PR comment posts on every pull request.