User: "audit https://github.com/someone/some-mcp-server"
  └─ TrueForge agent (mcp-vetter)
       ├─ clone_target          ── shallow-clones the GitHub URL onto the probe host
       ├─ read_target_manifest   ── declared tools & permission boundaries
       ├─ subagent: static_audit ── AST rules + Semgrep (SENT-001..007)
       ├─ subagent: full_audit   ── GPT review + Docker probes (SENT-008..011)
       ├─ Synthesizes verdict (HIGH/MEDIUM/LOW, OWASP Agentic Top 10 mapped)
       └─ ⏸ PAUSES before filing the GitHub security issue → human approves → files

# 1. Probe server (the security scanning engine)
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
uvicorn probe_server.server:mcp_app --host 127.0.0.1 --port 8000   # serves /mcp

# 2. TrueForge (separate terminal; needs Node 22+)
npx @truefoundry/trueforge@latest                                   # UI at http://localhost:8790

# 3. In the TrueForge UI:
#    Settings → Models      : configure a provider (API key)
#    Settings → Connectors  : Add MCP Server → http://127.0.0.1:8000/mcp
#                             (+ GitHub connector from the catalog, OAuth)
#    Settings → Sandbox     : optional (skills/code mode); cloning runs on the probe host
#    Create agent           : import deploy/agent-manifest.json via API, or compose in UI

# 4. Chat: "audit ./fixtures/vulnerable_server"

probe_server/    MCP server exposing the scanning engine as agent tools
fixtures/        Vulnerable + hardened reference MCP servers (from Sentinel, MIT)
deploy/          TrueForge agent manifest (agent spec via API)
docs/            PRD, architecture, week plan, setup guide