MCP setup (Claude Code, Cursor, Antigravity)¶
Claude Code (recommended)¶
Cursor and other MCP clients¶
npx trace-sec install-skill adds a trace server to .mcp.json and
.cursor/mcp.json and keeps any servers already listed there (--no-mcp
skips it). trace setup-mcp writes .mcp.json only, also preserving other
servers.
Manual bridge¶
Connect the agent to http://127.0.0.1:8765/sse.
Tools the agent gains¶
| Tool | Purpose |
|---|---|
trace_scan |
Full audit; returns counts, endpoints audited and choke points. profile: quick skips live probes; scope mirrors the CLI; non-local target values are refused |
trace_status |
Whether the repo is initialized, total findings, confirmed findings |
trace_findings |
Query findings; filter by minimum severity, page with limit / offset |
trace_explain |
Root cause, evidence, reproduction steps and remediation for a finding |
trace_attack_path |
APM hops for an endpoint |
trace_verify |
Prove the exploit vector is closed (FIXED / STILL_PRESENT / INCONCLUSIVE) |
trace_harness_task |
Export a SWE-bench-style task instance |
trace_eval_patch |
Score a proposed patch, optionally rolling it back |
Scans skip file formats TRACE cannot analyze; see What TRACE scans. Pair the tools with the agent skills so the agent also knows how to read the results.
See also: Security harness skill.