1. Connect your coding agent
Open the installed desktop client, review the server configuration, then complete its Firedrill browser sign-in. VS Code connects through Copilot; the Codex extension uses Codex’s separate MCP settings. These buttons do not install skills, authorize access, or run a simulation.
- Codex
- Claude Code
Inspect this agent’s real entry point and dependency connections. Prepare a Firedrill recipe with a few useful scenarios. Keep my agent running here, route any external side effects to synthetic dependencies, and do not replace its logic with expected answers. Start with one trial, then explain the outcomes and evidence.For a command or HTTP entry point, your coding agent writes a recipe and a small adapter if needed. For a callable function, use the TypeScript or Python SDK. The first-run guide includes a complete command-adapter example.
2. Prepare local execution
The MCP can prepare and inspect work, but cannot start a process on your laptop. For a command or HTTP simulation, initialize from your application repository with Node.js 22 or newer. No global installation is needed:npx firedrill-run init --skip-install to select scope without these additions.
Interactive init then offers its own browser sign-in if needed. Use the same
organization and project as the MCP connection, then select its environment.
Every created project has a Default environment. Init saves public IDs in
firedrill.config.json, not credentials. Do not copy the MCP host’s OAuth token
to the CLI.
For repeated use, optionally install npm install --global firedrill-run.
You can then use the shorter firedrill commands shown in the reference guides.
Python applications can use the native Python SDK without Node.
CI uses a scoped machine credential from
its secret store and never opens a browser.
3. Review and run
Review the recipe prepared for your application, then run:firedrill/; you can also supply an exact path, such as
npx firedrill-run simulate firedrill.recipe.json.
Add --choose to select individual scenarios with Space.
Simulate shows its resource estimate, invokes your real entry point and saves its
runs. Preserve the returned operation ID and recovery command if interrupted.
For checks without execution, replace simulate with doctor or preview.
The CLI returns an authenticated report link. Open it to see actual input/output, execution, optional checks and quality scores, activity, and captured files.
4. Grow the simulation
- Add synthetic Tools only for dependencies your agent uses.
- Add datasets for different starting conditions.
- Vary scenarios, repetitions, and configurations.
- Add semantic scoring or exact checks when you want assessment.
- Save the same runner in CI.