cancel_job
ChatGPTCancel a pending or running Comfy Cloud job by its prompt_id. Has no effect on a job that has already finished, and never deletes outputs already produced. Returns plain text saying which happened: a real cancellation, an already-finished no-op, or an unconfirmed cancel — relay that, do not assume the run stopped.
estimate_credits
ChatGPTEstimate the credit cost of a Comfy Cloud template or workflow BEFORE running it. Read-only: nothing is submitted or spent. Pass exactly one of template_name (a template name; fetched server-side) or workflow. Prices come from the bundled node-catalog declarations and can lag live prices; GPU/queue time and storage are excluded. Full caveats ride on the result.
get_feedback_survey_link
ChatGPTReturn the Comfy Cloud feedback survey link. Use when the user offers or asks about feedback. Returns a short thank-you plus the survey URL — relay both verbatim. Records nothing beyond the fact that the link was shown; all feedback is given in the survey.
get_job_status
ChatGPTCheck the execution status of a submitted Comfy Cloud job by its prompt_id — a single point-in-time snapshot. Returns status: in_progress, completed, error, or cancelled, plus raw_status from the API (e.g. submitted, queued_limited, executing). On error/cancelled, billed is false if the job never started executing, else "unknown".
get_output
ChatGPTRetrieve the outputs (images, videos, audio) of a completed Comfy Cloud job. Does not wait. Returns a no-auth download link (normally a /api/s/<id> shortlink) per file plus a browser-renderable inline URL where available; surface the link VERBATIM and never resolve it to the underlying signed URL. For PNG outputs it also reports pixel width/height. inline_urls: false skips the inline URLs.
get_saved_workflow
ChatGPTGet one saved workflow from the user's Comfy Cloud workspace by workflow_id (UUID or https://cloud.comfy.org/#<uuid> link) or filename. Read-only. Returns a node summary, the workflow JSON, its detected format, and a customizable_inputs breakdown of the per-run parameters (prompts, seeds, models, images) with the node id and input name each one is set through. summary_only: true returns a compact response without the full JSON.
list_saved_workflows
ChatGPTList the workflows saved in the user's Comfy Cloud workspace (cloud UI, comfy-cli, and MCP-saved). Each entry has a filename plus workflow_id and version (both null for legacy file-only entries); filename_ambiguous: true means that filename does not resolve to that row — use workflow_id. Paginated (default 20): while has_more is true, re-call with offset: next_offset to get the rest.
partner_generate
ChatGPTGenerate an image, video or audio clip with a partner provider model (Grok, Gemini, Flux, OpenAI, Kling, ByteDance, ElevenLabs, …) from a text prompt and optional input media. Diagrams/infographics are images. Pass the provider model slug (e.g. bfl/flux-2-pro); an unknown slug returns the closest real slugs. Returns {prompt_id, status} for a job whose result is saved to the library, or a direct download link for edits that cannot be saved. Spend-gated: call it WITHOUT confirm first — if it runs, done; if it returns a credit-spend prompt, relay it and re-call with confirm: true only on the user's explicit yes.
run_saved_workflow
ChatGPTRun a workflow saved in the user's Comfy Cloud workspace, identified by filename or workflow_id. The server fetches it, converts it and submits it; it never accepts raw workflow JSON. Only save-format (graph) workflows run here: a saved workflow whose format is api cannot be run on this connection. Optional per-run parameters: input_overrides ({nodeId: {inputName: value}}, the ids from the saved workflow's customizable_inputs) changes prompt, seed and similar inputs; slot_overrides sets subgraph-exposed slots by address (e.g. "57.text"). Returns a prompt_id. Spend-gated when the workflow uses paid API nodes: confirm the user agrees to spend credits, then re-call with confirm: true. include_workflow: true also echoes the converted workflow JSON.
run_template
ChatGPTFetch a Comfy Cloud template by name and submit it without loading the workflow JSON into context. Optional prompt sets its primary text input and seed sets its primary seed input — omit either to keep the template's own default; if the template has neither, passing one is rejected rather than silently ignored. Returns a prompt_id to poll (or one-shot, below). Spend-gated when the template uses paid API nodes: call it WITHOUT confirm first — if it runs, done; if it returns a prompt, relay it and re-call with confirm: true on an explicit yes. One-shot: wait_for_output: true (with client_os) waits inline (~25 s) and returns the finished output links, or {status: "in_progress", timed_out: true} if still running.
search_models
ChatGPTSearch the Comfy Cloud model catalog by text query, type, base model, or source. Returns lean fields; detail:"full" adds source URLs. Partner/API models are flagged source:"partner" with type = their modality (image/video/audio/…); filter source: "partner" for the partner list, and a tiered partner entry carries tiers. The user's own imported models are flagged owned_by_user; owned: true filters to them. Every page reports metadata.total and metadata.returned; while offset + returned < total, re-call with offset = offset + returned.
search_templates
ChatGPTSearch the comfy.org workflow template library (text-to-image, image-to-video, style transfer, audio, 3D, …) by text query and filters. Read-only. Returns each template's name, title, description, media type and tags; detail:"full" adds every field. Results are paginated and state their own truncation.
upload_file
ChatGPTUpload a local image/audio/video file (.jpg/.jpeg/.png/.webp/.gif/.mp3/.wav/.flac/.ogg/.mp4/.webm/.mov only) via a credential-free, single-use, short-TTL upload URL on the Comfy Cloud API host. Returns a PUT command for the user to run in their own shell; on success it prints a JSON body with the input name, which is then used as an input image/audio/video for a generation. If the command fails, call this tool again for a fresh URL. File bytes are never sent as a tool argument.
wait_for_job
ChatGPTBlock until a submitted Comfy Cloud job reaches a terminal state (completed, error, or cancelled), collapsing repeated status checks into one call. Pass max_wait_seconds sized to the job (video / multi-node templates: 300); omitting it holds 25 s. If it returns { timed_out: true }, call it again on the same prompt_id; tell the user if it keeps timing out or returns stalled: true.