Build a ComfyUI workflow with a coding agent
Inspect dependencies, run one known workflow, and preserve a recipe you can repeat.
Outcome and prerequisites
Produce one image from a known workflow and save its graph, input settings, model/node dependencies, and job identity. You need a coding agent with access to your project and workflow files, a running ComfyUI installation, compatible model files, and permission to use its compute. Start with Understand image models. No GPU run is implied by an offline graph check.
The workflow assistant structures the procedure, the workflow inspector checks graph structure, and the job client submits and collects a local job. Download the workflow dependency worksheet for the API starter and blank/filled dependency records.
1. Choose the connection before credentials
A self-hosted ComfyUI server exposes local HTTP routes such as /object_info, /prompt, /history/{prompt_id}, and /view; use the address of the actual installation. 127.0.0.1 refers to the machine running the caller, so a hosted web server cannot reach your laptop through its own loopback address. See self-hosted routes.
A hosted service has its own authentication and endpoint contract. Use that provider's current documentation; do not add an invented API-key header to the local runner. Comfy's old v1 Cloud API is deprecated in favor of its newer API; this local Vibe is not a Cloud API adapter. See Cloud API status.
Vibes for this guide
Use the materials that help with your next step.