Skip to main content
The Control MCP client helps a coding agent discover projects, selected Tools, saved tests, results and evidence metadata. It is different from a synthetic Tool’s MCP endpoint: Control MCP manages Firedrill resources, while your evaluated agent uses a scoped connection inside one case.

​
Configure the client

Configure your coding agent’s MCP client to launch:
Provide FIREDRILL_CREDENTIAL securely in that process’s environment using your MCP client’s secret settings. Do not commit a token into this JSON or assume every MCP client expands shell variables in configuration. FIREDRILL_API_URL defaults to https://api.firedrill.run. Use a project credential with only the permissions the coding agent needs. Normal permission checks remain enforced by Firedrill; installing the client does not grant more access.

​
Read-only by default

Control MCP exposes explicit, bounded operations rather than a generic HTTP call. It can read resource identities, Tool contracts, saved tests, retained verdicts, comparison summaries and evidence metadata. It does not download original artifacts, inspect hidden live state or forward authored Tool-call payloads. It cannot issue a world credential or run the customer agent. Connection guidance prints a read-only discovery command and a separate explicit issuance command for the operator to run locally. Keep any resulting scoped credential out of chat and unrelated processes.

​
Optional bounded mutations

To expose supported mutations, add --allow-mutations to the executable’s arguments. Mutations still require their documented exact inputs, permissions, confirmation and idempotency keys. Do not enable them simply to read reports. Starting a session is not starting the evaluated agent or producing a pass. Your coding agent writes reviewed test and adapter files in your repository, then uses the CLI or native SDK to run them.

​
Inspect simulations

The client reads saved simulation batches, immutable plans, bounded case pages and native attempts. The opt-in preview operation reviews an exact selection without invoking agents, launching cases or reserving capacity. The comparison operation accepts up to eight explicit retained batches from one authorized project. It reports native outcomes and optional measurement coverage, latency, usage and reported/estimated cost. It is advertised only when the configured client supports comparison; an older adapter is not silently presented as a working capability. Use model benchmarks to execute the same cases against actual customer model configurations. MCP report inspection does not execute or judge those configurations.