Skip to main content
A command or HTTP target can participate in a simulation without a Firedrill SDK inside the agent. Your customer-owned runner executes the command or contacts the endpoint. The operator MCP can prepare the batch but cannot launch an arbitrary local process on its own.

​
Command target

Declare literal argv in the recipe:
The runner writes a JSON invocation envelope to stdin. Read input for the task, configuration for selected parameters, and tools for the current run’s scoped connections. Return the real agent output on stdout. Valid JSON remains structured output. Write diagnostics to stderr and keep secrets out of both streams. See the complete command-adapter guide.

​
HTTP target

Use the target schema to declare the existing endpoint and request configuration:
The HTTP target receives the same task envelope as its request body. Firedrill operator credentials are not added as the target’s authorization. Supply only the target authentication your application needs through the supported secret mechanism described by the schema and SDK guide. Reachability depends on the runner’s network. A local endpoint is not reachable from every CI machine.

​
Trust and recovery

Only connect endpoints and executables you control and intend to exercise. A command is not a hostile-code sandbox; an HTTP timeout does not prove its side effects stopped. Preserve the execution operation ID after uncertainty. Do not rerun a target merely to recreate a missing terminal response or report.