Scan
agentctl scan ./workspace inventories agent entrypoints, registry entries, model references, tools and MCP policy sources - read-only, with test/example paths excluded by default.
Step 2: Verify
Know which agents actually run in your code.
Agent Control Plane scans an approved local directory in read-only mode and builds an inventory: agent entrypoints, model and provider references, tools, MCP servers and identities - then flags deterministic risks with file and line evidence.
Written in Go. SARIF 2.1.0 output plugs straight into GitHub Code Scanning and CI gates.
The pain it solves
Agent frameworks make it easy to add a capability and hard to notice one. MCP servers multiply that problem: every config file is a new permission grant. Security teams cannot review what they cannot list.
Common triggers
- Nobody can list every agent and MCP server in the repo.
- A security review starts and there is no inventory to review.
- An MCP config grants more than anyone remembers approving.
- Runtime behavior drifts from the last known baseline.
How it works
Static scan first, runtime correlation second, CI enforcement third.
Audit runtime
agentctl runtime-audit correlates observed agent activity (JSONL, OpenTelemetry, API gateway events) with the static inventory. Metadata-only input: secrets and payloads are rejected or omitted.
Diff and gate
agentctl runtime-diff compares snapshots and exports text, JSON, CSV or HTML. In CI, --fail-on high turns findings into a merge gate, with reviewed baselines and expiring suppressions.
Review the evidence
Every finding points to a file and line. SARIF output renders natively in GitHub Code Scanning and other security tooling.
Next step
Verified is not proven.
An independent review turns your inventory and findings into a report that clients, auditors and procurement teams accept.