Skip to main content
Glama

comfyops-mcp — Local Generative AI Engine

Wraps ComfyUI so you can generate images, video, and upscales from a prompt — without ever touching the node editor. Pick a curated workflow, type what you want, get a PNG or MP4.

FastMCP 3.4+ | 6 portmanteau tools | 29 curated workflows | ComfyUI :11086 | Backend :11087 | Dashboard :11088

Comfy for the Perplexed

What's ComfyUI? It's the engine room of local AI image generation. You connect boxes labelled "KSampler", "CLIP Text Encode", and "VAE Decode" with wires, like a Moog synthesizer made by someone who hates knobs. It's incredibly powerful and completely impenetrable.

What's this? comfyops-mcp is a butler for ComfyUI. You tell it what you want — "a cat surfing on a pizza slice" — and it picks the right boxes, wires them up, and hands you the result. No node editor, no patch cables, no "what's a CFG scale?"

What can it do?

  • Turn text into images — verified: FLUX.2 klein 4B (seed-reproducible) and SD 1.5; curated: SDXL, FLUX.1 family, Z-Image, PixArt, Lumina2

  • Turn text or photos into short videos — curated (Wan 2.1/2.2, Hunyuan, LTX, Mochi, Cosmos) — not yet end-to-end verified

  • Upscale / restore old images — curated (ESRGAN, SUPIR) — not yet end-to-end verified

  • Inpaint and edit (SD1.5 inpaint/img2img; FLUX fill curated)

  • Browse and catalog workflows from the community

  • Remember every generation with its seed so you can reproduce it later

Who is it for?

  • You, who just wants pictures — prompt in, PNG out

  • You, who has an RTX 4090 gathering dust — put it to work

  • You, who already uses ComfyUI — now your AI assistant can drive it

  • You, who saw a cool AI image online and wants to make one

If you already know ComfyUI: comfyops registers curated workflows from workflows/ as MCP tools. You can add your own with comfy_workflows/register. The workflows dir is yours — export from the node editor, drop in, done.

Related MCP server: ComfyUI-MCP-Server-Python

Preview

ComfyUI Node Editor (what we wrap)

comfyops Dashboard (what you see)

ComfyUI

Dashboard

Screenshots pending — run just screenshots once the dashboard is live.

What You Can Do

"Generate a 1024x1024 image of a cat surfing on a pizza slice"
"Upscale my archive of 2024 SD generations to 4K"
"Turn this photo into a cinematic video" (Wan 2.2 i2v)
"Make 10 variations of this character concept"

Quick Install

git clone https://github.com/sandraschi/comfyops-mcp
cd comfyops-mcp
uv sync
.\start.ps1

See docs/ONBOARDING.md for first-run (ComfyUI + models), and INSTALL.md for Claude Desktop config.

Models from Civitai: use civitai-mcp to search/download into the ComfyUI models tree; this repo generates.

Features

  • Prompt-to-image — verified: FLUX.2 klein 4B fp8 (official ComfyUI wiring, ~1-2 min per 1024x1024 on a 4090, seed-reproducible) and SD 1.5

  • Prompt-to-video / upscale / inpaint: curated workflows (Wan 2.1/2.2, Hunyuan, LTX, Mochi, ESRGAN, SUPIR) — pending end-to-end verification

  • Model downloads: comfy_models/download streams from Hugging Face with optional sha256 verification and Bearer auth for gated repos

  • VRAM guard: Checks free GPU memory before queueing — no OOM crashes; models are unloaded after each run so the reading stays honest

  • Seed control: Same seed + same workflow = identical output, every time (verified)

  • Generation library & gallery: SQLite history with sorting, filters, batch delete/export (CSV/JSON), and crossconnects (same workflow/model)

  • Agentic workflow: Multi-step generation planning via MCP sampling (SEP-1577)

  • Prefab cards: Rich in-chat status and generation cards

How It Runs

Mode

ComfyUI required?

When

MCP tools (stdio/HTTP)

Yes

Generation via Claude Desktop, Cursor, opencode

Dashboard (Vite)

No

Browse workflows, models, library — generation needs ComfyUI

ComfyUI sidecar

Optional

start.ps1 can launch ComfyUI automatically if COMFYOPS_COMFYUI_DIR is set

Workflow Depot

comfyops ships with 27 curated workflows (t2i, i2v, t2v, inpaint, edit, upscale, restore). Use comfy_workflows/register to add your own exports from the ComfyUI node editor.

Tools

Tool

Ops

comfy_generate

image, video, upscale, inpaint, edit

comfy_comic

generate_panel, assemble_strip (two-pass hybrid, WS-first)

comfy_workflows

list, get, validate, register

comfy_models

list_installed, download, check_vram, health

comfy_library

recent, search, record

comfy_agentic_assist

Multi-step via MCP sampling

Community

Available Tools

8 tools
comfy_agentic_assistA

Multi-step agentic generation via MCP sampling (SEP-1577).

Plans a generation campaign: describes the creative goal, selects appropriate workflow(s), generates, optionally vision-checks via local multimodal model, and retries with adjusted params (max 3).

Requires a host with MCP sampling. Falls back to a structured manual tool sequence when sampling is unavailable.

Return Format

{"success": bool, "agent_plan": str, "error": str}

Examples

comfy_agentic_assist(goal="Create a cinematic aerial shot of a cyberpunk city")

Inputs: goal (required). Returns: object

comfy_generateA

Generate image, video, or upscale via a curated ComfyUI workflow.

RATIONALE

Consolidated portmanteau: all generation modes share the same ComfyUI workflow submission and result polling pipeline. The operation discriminator selects the workflow type.

Return Format

{"success": bool, "prompt_id": str, "outputs": [...], "seed": int, "message": str}

Examples

comfy_generate(operation="image", workflow_id="flux-klein-t2i",
               prompt="a cat surfing on a pizza slice", seed=42)
comfy_generate(operation="upscale", workflow_id="esrgan-upscale",
               prompt="", image_input="<base64>")

Inputs: prompt (required), operation (required), workflow_id (required), seed, +4 more. Returns: object

comfy_libraryC

Browse past generations and record new ones.

Return Format

{"success": bool, "generations": [...], "message": str}

Examples

comfy_library(operation="recent", limit=10)
comfy_library(operation="search", query="cat surfing")

Inputs: operation (required), limit, model, query, +5 more. Returns: object

comfy_modelsA

Manage local models, download from Hugging Face (hash-verified), and check GPU VRAM.

Return Format

{"success": bool, "models": [...], "download": {...}, "vram": {...}, "health": {...}}

Examples

comfy_models(operation="list_installed")
comfy_models(operation="check_vram", model_vram_gb=6.0)
comfy_models(operation="download", hf_repo="org/model", filename="m.safetensors",
             target="diffusion_models", sha256="<hash>")
comfy_models(operation="health")

Inputs: operation (required), sha256, target, hf_repo, +3 more. Returns: object

comfy_nodesA

Install and resolve ComfyUI custom nodes via ComfyUI-Manager.

Maps missing node class_types → Manager packages (registry/GitHub catalog), installs via cm-cli or Manager REST queue, restarts ComfyUI, re-validates.

Return Format

{"success": bool, "packages": [...], "mapped": {...}, "message": str}

Examples

comfy_nodes(operation="status")
comfy_nodes(operation="ensure", workflow_id="my-complex-wf")
comfy_nodes(operation="install", packages="ComfyUI-Impact-Pack,ComfyUI-VideoHelperSuite")
comfy_nodes(operation="resolve", missing_types="ImpactInt,VHS_LoadVideo")

Inputs: operation (required), git_url, packages, workflow_id, +3 more. Returns: object

comfy_workflowsB

Manage curated ComfyUI workflow definitions.

Return Format

{"success": bool, "workflows": [...], "workflow": {...}, "sources": [...], "message": str}

Examples

comfy_workflows(operation="list")
comfy_workflows(operation="search", query="flux", tags="video")
comfy_workflows(operation="discover")
comfy_workflows(operation="register", workflow_id="my-wf",
                workflow_json='{"3":{"class_type":"KSampler",...}}',
                tags="t2i,portrait", source_url="https://civitai.com/...")

Inputs: operation (required), name, tags, query, +4 more. Returns: object

show_comfyops_status_cardB

Show ComfyUI health and generation status as a rich in-chat card.

Return Format

Prefab UI card with system status, VRAM, workflow count, model count

Inputs: no named parameters. Returns: object

show_generation_cardC

Display a single generation result as a Prefab card.

Return Format

Rich card with prompt, seed, workflow, and output files

Inputs: seed (required), prompt (required), workflow_id (required), outputs. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.5/5.0

    Scored across 8 tools

    Disambiguation4/5

    Most tools map cleanly to a distinct resource or action: generation, workflow definitions, models, nodes, generation history, and UI cards. The main boundary overlaps are comfy_agentic_assist vs comfy_generate and comfy_models health vs show_comfyops_status_card, but the descriptions and operation parameters give enough guidance for an agent to choose correctly.

    Naming Consistency4/5

    The core tools share a predictable comfy_ prefix and snake_case style, with each comfy_<domain> tool using an operation parameter to select actions. The show_*_card presentation tools and comfy_agentic_assist deviate from the resource-manager pattern, but the naming remains readable and mostly consistent.

    Tool Count5/5

    Eight tools is a well-scoped size for a ComfyUI operations server; each tool covers a meaningful area such as generation, workflow registry, model management, node resolution, generation history, agentic orchestration, and UI display. The operation-based design keeps the tool count from ballooning while still exposing the needed functionality.

    Completeness4/5

    The core generation lifecycle is covered end-to-end: discover/register workflows, resolve nodes, manage models, generate, browse history, and display results. Gaps such as deleting/uninstalling workflows/models/nodes and an explicit library record operation keep it from being fully complete, but agents can work around these for typical generation tasks.

    Maintenance

    ActivityActive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers