The outcome

Expose automation without granting a leaked key broad team access or letting model-selected tools perform uncontrolled actions.

Step by step

A workflow you can repeat.

  1. 01

    Map inputs, outputs, every node and branch, team resources, external systems, permissions, side effects, idempotency, human approvals, timeouts, concurrency, quotas, SLO, and rollback.

  2. 02

    Separate retrieval, classification, transformation, and action nodes, use schemas and explicit conditions, restrict HTTP destinations, sandbox code, minimize variables, and keep model-chosen tools read-only.

  3. 03

    Create a dedicated least-privilege member and API key, set expiration and quota, store it server-side, pass the explicit app ID, and never assume an app-publishing key is app-scoped.

  4. 04

    Test all branches plus injection, malformed variables, unauthorized resources, SSRF, duplicate requests, loops, parallel limits, timeouts, provider outage, key expiry, denial, and compensation.

  5. 05

    Canary traffic, log redacted node outcomes, monitor errors, latency and spend, rotate keys, retain the prior workflow export, and verify revoked keys, deleted runs, and retired resources.

Working standard

What good use looks like.

  • Assume member keys reach team resources.
  • Restrict HTTP and code nodes.
  • Keep consequential release human-controlled.

Official references

Check the current product documentation.