Skip to main content
Use synthetic Tools when your agent interacts with external services. You can start them manually for exploration or let a simulation allocate isolated copies automatically. A Tool setup does not run your agent. Your agent continues in your own application or runner.

​
Choose Tools interactively

After initializing your project:
The CLI guides you through:
  1. The Tool catalog or your saved private builds.
  2. One or more Tools: arrow keys move, Space selects, Enter continues.
  3. Allowed operations, starting data, and independent copies.
  4. Instance names and a lifetime.
  5. A confirmation before allocation.
You can select the same Tool more than once with distinct names. Each instance has separate data. Data choices are explicit: catalog starter data, empty state, or a saved dataset for its compatible build. The default manual lifetime is 30 minutes. Choose a longer lifetime in the prompt or use --ttl-seconds with an interactive start.
This requests four hours, subject to the service’s current resource envelope. Extend a current instance before it expires.

​
Inspect readiness and history

Replace INSTANCE_ID with the real identity returned by startup or the instance list. A pending or uncertain startup is not a ready connection. A partial setup retains its successful instances rather than blocking unrelated setups. Use the saved recovery command to reconcile the same setup:
The original key and private journal preserve the selected releases and known instances. Do not start another setup just to recover an uncertain reply.

​
Use a saved plan from code or CI

A noninteractive start uses a complete plan:
Each plan item declares a unique alias, exact catalog release or private build, starting-data choice, allowed operations, seed, and lifetime. Your coding agent can obtain those real identities with firedrill tools catalog --json and firedrill tools builds --json, then prepare the plan using the SDK contract.

​
Use Tools in a simulation

You do not need to start manual instances first. Pin the returned build and dataset in the recipe’s services bindings; the runner allocates separate instances for each independent case and closes them after execution. See connect your agent. The web app also supports manual Tool setup and shows instances created from CLI or SDK workflows. Available browser apps show the instance’s actual selected data.