Skip to main content
Executes a handler(input) function in a secure Firecracker sandbox and returns markdown plus structured data. Display name “Python Code”. Off by default — enable it in the agent’s Tools panel. This is not how you render HTML in the task feed.

​
When to use it (vs app_tool vs custom_interface vs generate_file)

handler(input) must return {'markdown': str, 'data': any}. That markdown is what the feed shows — not HTML. There is no flag to “render as HTML”. If the user wants a scorecard or interactive widget in the feed, use app_tool (interactive) or custom_interface (display-only).

​
Authentication and enablement

No OAuth. Runs in the platform sandbox (HTTP calls capped, time/memory limited). Off by default.

​
Inputs

  • python_code (required for the base tool; injected from bindings on a configured tool): a handler(input) function.
  • input (required): extra parameters for the handler. Never pass "input": {} if the instructions mention parameters. Task variables (VAR_1, VAR_2, …) are auto-injected — read them as input['VAR_1']; do not put them in input yourself.
Python syntax: None / True / False, not JSON null / true / false.

​
Output

Structured result with output (the handler return), prints (stdout), and optional artifacts from save_artifact(name, content). The feed renders output.markdown.

​
Limits and side effects

  • Sandbox: bounded CPU, memory, and outbound HTTP.
  • save_artifact uploads files the user can download; call it once per file.
  • Configured-tool code is stored in bindings and not shown as a per-run input on later runs.

​
Expected errors

  • Empty input when parameters were required — the handler then has nothing to work with.
  • Handler does not return {'markdown', 'data'} — treat as a contract violation and fix the return.
  • Sandbox timeout / HTTP budget exceeded.