Skip to main content
Firedrill is a managed simulation platform for AI agents. Ask your coding agent to connect your application’s existing entry point, prepare scenarios, and run them. Your application stays in your repository, workstation, or CI runner. Use simulations to explore possible outcomes, demonstrate a workflow, compare model configurations, verify expectations, or find failures. Synthetic Tools and quality scoring are optional.

​
Start from your repository

Approve Firedrill in the browser. The remote MCP needs no local Firedrill package. The coding-agent guide has native skills/plugin commands, Cursor configuration and Kimi’s setup flow. Instructions are separate from the connection; host permissions and browser authorization remain explicit. Then ask your coding agent:
Connect this application’s real agent to Firedrill. Prepare a small set of representative scenarios, including interesting boundaries. Run them, show the recorded outcomes and evidence, and explain what remains unknown. Keep production systems untouched.
Your agent uses the CLI or SDK to execute your application in your own environment. The remote MCP prepares work and reads its retained results; it does not start a process on your machine.

Run your first simulation

Authenticate, connect your application, and open its recorded result.

Add synthetic Tools

Give your agent controlled dependencies and explicit starting data.

Use TypeScript or Python

Call your existing agent from a small native SDK adapter.

Run in CI

Reuse the same scenarios on PRs, pushes, schedules, or manual runs.

​
Understand what happened

Looking for a similar application? Browse AI agent examples by framework and stack, including LangGraph, Pydantic AI, Google ADK, Mastra, CopilotKit, and browser agents. Each execution becomes a run. A batch groups scenarios, configurations, and repeated trials. Reports keep four questions separate:
  • Did the agent execution complete?
  • What did it actually do or return?
  • Did optional exact checks or quality criteria hold?
  • What did the optional CI policy decide?
Different wording can be equally correct. Use quality scoring for meaning and behavior; use exact checks for hard requirements. A completed run is not automatically a pass, and an evaluation problem is not an agent failure. The web app shows the same saved runs, scenarios, evidence, and Tool instances as the CLI and SDKs. It does not need to launch your local agent.