Deploy your connected application

Ship the generated helper, connection configuration and runtime secrets with the Python application that makes model calls.

Connect to the exact HTTPS source URL returned in Data. Your application keeps using its OpenAI or Anthropic Python client library and provider credential. See Quickstart to create the connection.

Application files

Deploy the generated reasonblocks_setup.py and its adjacent .reasonblocks/config.json with the application. Include the hidden config directory explicitly in wheel package data or image copy rules. The generated .reasonblocks/SETUP.md is optional at runtime.

Install rbtrace==1.2.1 from PyPI into the application’s Python environment and record the version in its dependency manifest. That release contains the CLI and the helper’s automatic run labelling.

Use an importable helper location appropriate to your package. Verify the real entrypoint from a working directory outside the repository. See helper placement.

Runtime secrets

Supply REASONBLOCKS_CAPTURE_KEY through the deployment’s secret environment. The existing RB_CAPTURE_KEY alias is accepted; if both names are set, their values must agree. Keep the provider credential in its normal secret location. Do not put either key in generated files or the image.

The capture service receives model requests and the provider credential needed to forward them. Review the workflow content you send. Capture keys expire after seven days; rotating a key in Data invalidates the previous key, so update connected applications as part of that change.

Tasks and verification

Wire the helper into the process that makes the actual model calls. Create one run ID per task and reuse its headers for every step, including in workers that handle many tasks. See Integrating your agent.

Run doctor in the deployed Python environment with the configured helper path. It is a local check, not a remote credential or capture test. Confirm source records after a real workflow before claiming the deployment is capturing correctly. Normal provider charges apply to those calls.

Training environment

Capturing calls does not require a sandbox. Full-agent training later needs a resettable test copy of your tools and data plus an outcome evaluator. The adapter can connect an existing test environment; this does not inherently require a new customer-hosted server. Follow Train your complete agent for the application-specific connection and release requirements.