MCP server for interacting with ComfyUI, enabling image/video/audio generation, workflow management, model discovery, and job tracking
The server has 29 well-named tools with mostly present descriptions and schemas, but exhibits inconsistent quality across tools. Naming convention is strong (verb_noun pattern, action-oriented). Many descriptions are good length (50-250 chars), but some lack critical guidance on dependencies, constraints, and error recovery. Parameter schemas are present and mostly typed, but several tools lack descriptions for their parameters. Output schemas are not documented anywhere, LLMs cannot predict return structures. Error handling is minimal; tools rarely guide recovery. The server follows basic MCP patterns but misses several production-grade practices around composition, pagination, and result capping.
Build a ComfyUI node with validated inputs
Cancel a generation job by task ID or prompt ID
Delete a custom saved workflow template
Download a font file for use in ComfyUI text rendering (Google Fonts, system fonts, or custom URLs)
Find ComfyUI nodes by their type or capability
Get download URL for a model from HuggingFace or Civitai
Fetch an example workflow (extracts embedded JSON from documentation images). Returns ready-to-use workflow JSON that can be passed directly to run_workflow. Use this as a starting point for creating custom workflows.
Get generation history for a specific prompt/task
No output schemas documented for any tool. LLMs cannot predict what fields to expect in responses, forcing them to reason about return structures and plan downstream calls blindly. This violates pattern:tool and pattern:tool-description.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 53 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 46 | - | v1 |
Get a generated image by filename, returns it as base64 or downloadable
Get installation instructions for ComfyUI (recommends desktop app for most users)
Get the status of an async generation job
Get guidance on downloading and installing models for ComfyUI
Get detailed information about a specific ComfyUI node including its inputs, outputs, and options
Get the current generation queue and execution status
Get the current status of ComfyUI connection and installation
Get a custom saved workflow template by name
Interrupt the currently executing generation
List official ComfyUI example workflows from the documentation. RECOMMENDED: Always check examples first before building custom workflows - they provide tested, working templates for common use cases like txt2img, img2img, ControlNet, LoRA, regional prompting, and more.
List available fonts for ComfyUI text rendering
List all generation jobs with optional status filter
List all available models in ComfyUI (checkpoints, LoRAs, VAEs, etc.)
List all available ComfyUI nodes with their descriptions
Get a recommended ComfyUI workflow template based on task description. Returns a ready-to-use workflow JSON that matches the requested task.
Render an SVG to PNG image using ComfyUI's SVG rendering node
Run a ComfyUI workflow synchronously and wait for completion (returns timeout error for workflows taking > 5 minutes)
Queue a workflow for asynchronous execution. Returns immediately with a task ID. Use get_job_status to poll for completion.
Save a workflow as a reusable custom template
Search custom saved workflow templates by name or tag
Validate a ComfyUI workflow JSON for correctness before execution
No pagination support on list_* tools (list_models, list_nodes, list_examples, list_jobs, list_fonts). Large result sets will blow context windows and lack nextpage/cursor guidance. Violates pattern:paginated-result.
Minimal error handling guidance. No tools include recovery instructions or actionable error messages. E.g., 'workflow validation failed' provides no hint for the LLM on what to fix or which tools to try next. Violates pattern:recovery-guide.
Many descriptive list_* and get_* tools lack clarity on when to use them versus similar tools. E.g., list_models vs list_nodes, when does the LLM choose one? list_examples vs recommend_workflow, what's the difference? Violates pattern:tool-description.
Parameter descriptions are sparse or generic on several tools. E.g., get_download_url's 'modelName' lacks guidance on format (HF model ID? Civitai URL? Plain name?). No validation rules stated. Violates pattern:tool-description and review:param-validation-rules.
run_workflow and run_workflow_async accept free-form 'workflow' objects with no documented structure. What fields are required? What are optional? How does the server validate them? Violates pattern:constrained-input.
No idempotency guidance on write operations. run_workflow, save_template, download_font, render_svg could have side effects on retry. If the agent retries on network failure, will the operation be duplicated? Violates pattern:idempotent-operation.
Destructive tools (delete_template, cancel_job, interrupt) lack confirmation/dry-run patterns. An agent could accidentally delete a template or cancel a critical job without user confirmation. Violates pattern:confirmation-request.
get_image returns base64 or downloadable URLs, but no output schema documented. How large can returned images be? What format? Should the LLM expect base64 inline or a URL? Violates pattern:tool-description and pattern:response-shaper.
run_workflow description says 'returns timeout error for workflows taking > 5 minutes' but no guidance on how to retry async, what the timeout error looks like, or when to call run_workflow_async instead. Violates pattern:recovery-guide.