Skip to main content
Glama

OpenDaemon ComfyUI MCP

Make AI videos on your own computer, by asking your AI assistant.

This is an MCP server. It lets any AI app that speaks MCP (Claude Desktop, Claude Code, Cursor, OpenDaemon and others) drive ComfyUI on your own graphics card. The assistant can:

  • render clips from a picture and an idea (image to video), or from the prompt alone (text to video), with settings picked for your card;

  • check its own work: it looks at frames of each clip, scores it and re-renders what went wrong (drift, melting, no motion...), up to a few tries;

  • mass-create: batches and variations wait in one queue, and your own renders always go first;

  • edit like a video editor: join, trim, speed, extend a clip, captions, voice-over, music, title and end cards, logo, reformat to 9:16 and export for TikTok, Reels, Shorts, YouTube...;

  • run your own ComfyUI workflows, check them before a slow run and install missing node packs (it asks you first).

Everything runs locally and costs nothing. Tested on a GTX 1050 Ti with 4 GB.

How it works

AI app (MCP client) --MCP--> this server --HTTP--> pipeline server :8190 --HTTP--> ComfyUI :8188 --> your GPU
                                                     (queue, editing, web page)

The AI app is the client. This server gives it tools (67 of them). When the model calls a tool, the server passes the job to a small pipeline server, which keeps one queue, talks to ComfyUI over its API and does the editing with ffmpeg. The pipeline server also serves a web page (http://127.0.0.1:8190) where you can watch the queue and render by hand. Renders started by the AI show up there too. If it isn't running, the MCP server starts it.

This project talks to ComfyUI only over its API and never bundles it.

Related MCP server: CutPilot

What you need

  • Windows with an NVIDIA graphics card, 4 GB or more. Other systems may work but aren't tested yet.

  • ComfyUI, for example the portable build, plus the ComfyUI-GGUF node pack for the GGUF models.

  • Python 3.10 or newer.

  • A video model. Ask the assistant: "list the video models" and "download the one that fits my card". Models come straight from their makers' Hugging Face pages and keep their own licences. Read them before commercial use.

Install

pip install git+https://github.com/open-daemon/opendaemon-comfyui-mcp

Or from a downloaded copy of this repo: pip install .

That installs two commands: opendaemon-comfyui-mcp (the MCP server) and opendaemon-comfyui-web (only the web page).

Connect your AI app

Set COMFYUI_DIR to the folder that holds ComfyUI's main.py, unless ComfyUI is in a usual place.

Claude Desktop: in claude_desktop_config.json:

{
  "mcpServers": {
    "comfyui": {
      "command": "opendaemon-comfyui-mcp",
      "env": { "COMFYUI_DIR": "C:/ComfyUI_windows_portable/ComfyUI" }
    }
  }
}

Claude Code:

claude mcp add comfyui --env COMFYUI_DIR=C:/ComfyUI_windows_portable/ComfyUI -- opendaemon-comfyui-mcp

Cursor: in .cursor/mcp.json, use the same block as Claude Desktop.

OpenDaemon: in opendeamon.json:

{ "mcp": { "servers": { "comfyui": {
  "command": "opendaemon-comfyui-mcp",
  "env": { "COMFYUI_DIR": "C:/ComfyUI_windows_portable/ComfyUI", "OPENDAEMON_COMFY_TOOLS": "core" } } } } }

Then ask, for example: "Make 3 short clips of the cat in C:/pics/cat.jpg and join them into one video."

Settings (environment variables)

Variable

What it does

Default

COMFYUI_DIR

ComfyUI's folder (the one with main.py)

looked for in the usual places

COMFYUI_URL

ComfyUI's address

http://127.0.0.1:8188

COMFYUI_PYTHON

the Python that runs ComfyUI, used to start it and to install node packs

the portable build's own

COMFY_MCP_DATA

where videos, your workflows and settings are kept

~/OpenDaemon-ComfyUI-MCP

OPENDAEMON_COMFY_TOOLS

full (67 tools), core (15 tools) or a comma list of tool names

full

GEMINI_API_KEY / NVIDIA_API_KEY

a free vision model key, used to write prompts and to judge clips

none: those two tools say so

FFMPEG

a specific ffmpeg

found by itself (one comes with this package)

PIPELINE_PORT / PIPELINE_URL

the pipeline server's port

8190

PIPELINE_AUTOSTART

0 = never start the pipeline server itself (for an app that runs it); PIPELINE_DOWN_HINT is then the message the model gets

1

Why core? An AI app sends every tool's description to the model with every message. All 67 tools cost about 7,000 tokens per message; the core 15 cost about 2,200. Small local models with a short context should use core.

Your own ComfyUI workflows

  1. In ComfyUI, open your workflow, then Workflow menu > Export (API). The normal save format can't be run over the API.

  2. Put the file in COMFY_MCP_DATA/workflows/, or ask the assistant to keep it (save_workflow).

  3. Ask for it by name. The assistant sees its inputs (list_workflows), checks it (check_workflow) and runs it (run_workflow).

Inputs are found automatically: the prompt and negative prompt, picture, video, audio, seed, steps, cfg, denoise, width, height, length and fps. To choose the inputs yourself, rename a node in ComfyUI to $name (its main input) or $name.field. Nodes with the same $name change together. A new random seed is used each run unless you give one.

There's an example in examples/workflows/ltx-image-to-video.json; it's copied into your workflows folder on first use.

The tools

Group

Tools

System

comfyui_status, start_comfyui, restart_comfyui, free_gpu_memory, get_logs, get_defaults, set_defaults

Models

list_models, download_model, download_status, cancel_download, search_models, check_model, add_model_file, remove_model

Pictures and prompts

upload_image, write_prompt, write_avoid_list, render_options

Render and queue

render, rerender, job_status, cancel, diagnose_job, estimate, render_batch, render_variations, queue, suggest_settings

Review and library

review_video, auto_fix, list_videos, video_info, delete_video

Editing

build_video, join_clips, trim_clip, extend_clip, change_speed, reverse_clip, loop_clip, boomerang, smooth_motion, upscale_clip, crop_zoom, color_adjust, fade, add_music, make_voice, add_voice, lipsync, add_captions, add_text, add_logo, title_card, end_card, reformat, export, thumbnail, make_gif, extract_audio, remove_audio

Your workflows

list_workflows, run_workflow, check_workflow, install_custom_node, save_workflow

The server also gives the model a short guide to prompting video models (the resource guide://prompting).

Safety

  • Anything that deletes, installs or downloads a lot asks you first: the tool returns needs_confirmation, and the assistant must ask you before calling it again with user_confirmed=true.

  • Node packs are programs written by other people. install_custom_node only takes GitHub links and always asks first.

  • The model list says how much graphics memory each model needs and which one is best for your card. Files found with search_models can only be downloaded if they fit your card.

  • Freeing GPU memory or restarting ComfyUI is refused while something renders.

Licence

MIT, see LICENSE. ComfyUI (GPL-3.0), the node packs and the models have their own licences.

Made by OpenDaemon.

Available Tools

67 tools
add_captionsA

Burned-in captions. text "auto" = timed to the voice line put on with add_voice (edge voices). Or your own text, spread over the clip ("|" forces a break). style: bold | impact | clean | box. position: bottom | lower | center | top. highlight = yellow text.

Inputs: video (required), text, style, position, +2 more.

add_model_fileA

Adds a model file that is already on this PC, where it is (a link, no copy). It is checked first. After this, render(model=) uses it.

Inputs: path (required), name.

add_musicA

Background music (a file path), looped or cut to the video's length, faded in and out. If the video has a voice, the music gets quieter while it speaks (duck).

Inputs: music (required), video (required), duck, volume.

add_textC

A title or any text on the video, from start to end seconds (default: the whole clip).

Inputs: text (required), video (required), box, end, +3 more.

add_voiceB

Puts a voice line (from make_voice, or any sound file) on a clip, starting at start seconds. If it is longer than the clip, the last frame is held until it ends.

Inputs: audio (required), video (required), start, volume, +1 more.

auto_fixA

The self-check loop for one video: a vision model judges it, the settings are fixed by the kind of problem (drift = less movement, no action = more movement, end melts = shorter, hands/melting = avoid list) and it renders again, at most tries times (1-3). Every try is kept; the answer names the best one. problems: if you already looked (review_video), say what you saw so the first judge call is skipped: drift, wrong_subject, no_action, too_much_motion, hands, melting, end_melts, talking, blurry. Returns a fix id; job_status(fix id) follows it.

Inputs: video (required), tries, priority, problems.

boomerangC

Forwards then backwards, like a boomerang. times = how many back-and-forths.

Inputs: video (required), times.

build_videoA

ONE call for a finished video. clips: video names in order, or {"video", "start", "end", "speed"} to cut/speed one. transition: cut | fade | crossfade. shape: "9:16" (Shorts/Reels/TikTok) | "1:1" | "4:5" | "16:9". voice: the words to say (or {"text", "voice": andrew|emma|ava|brian|es|pt..., "engine": edge|local, "sample"} or {"audio": file}). captions: "auto" (timed to the voice) or your own text. music: a music file path (or {"file", "volume"}); it gets quieter under the voice. title: text for a card at the start. end_card: text for the last card (true = "Follow for more"). export: shorts | reels | tiktok | facebook | youtube | x | square (right size and format to post). Only the final video stays in the gallery.

Inputs: clips (required), music, shape, title, +5 more.

cancelC

Stops a queued or rendering job.

Inputs: id (required).

cancel_downloadA

Stops the running model download. What was downloaded is kept, so starting it again resumes.

Inputs: no named parameters.

change_speedC

0.5 = half speed (slow motion), 2 = twice as fast. 0.1 to 8.

Inputs: speed (required), video (required).

check_modelA

Says if a model file can run here and why not if it can't, without loading it. A file on this PC: file="C:/path/model.gguf" or a file name in ComfyUI's models folder. A file on Hugging Face (from search_models): repo="owner/name" and file="path/in/repo.gguf".

Inputs: file, repo.

check_workflowA
Read-only

Before a slow run: finds missing node types (and which node pack has them), missing model files and wrong values in a saved workflow (name) or one you give (workflow, API format). Also lists its changeable inputs.

Inputs: name, workflow.

color_adjustC

brightness -1..1 (0 = same), contrast 0.5..2 (1 = same), saturation 0..3 (1 = same), warmth -1 cool .. 1 warm.

Inputs: video (required), warmth, contrast, brightness, +1 more.

comfyui_statusA
Read-only

Is ComfyUI running, which graphics card and how much memory it has, free disk space, and what is rendering.

Inputs: no named parameters.

crop_zoomA

A slow camera move on a finished clip. direction: in | out | left | right | up | down | none (a still crop). zoom 1.01-4. x, y (0-1) = the point to zoom on (0.5, 0.5 = the middle).

Inputs: video (required), x, y, zoom, +1 more.

delete_videoA
Destructive

Moves a video and its settings to the out/trash folder (not deleted for good). Ask the user first, then call with user_confirmed=true.

Inputs: video (required), user_confirmed.

diagnose_jobB
Read-only

Why a job failed, in plain words, with what to try and the ComfyUI log lines.

Inputs: id (required).

download_modelA

Downloads a model from our list (ids from list_models), with any helper files it needs. Big downloads: tell the user the size and call again with user_confirmed=true.

Inputs: model_id (required), user_confirmed.

download_statusC
Read-only

Progress of the running model download (one at a time).

Inputs: no named parameters.

end_cardC

A card at the END ("Follow for more", your logo...), faded in after the video.

Inputs: video (required), text, image, seconds.

estimateA
Read-only

How long a render (same settings as render) or a whole batch (items = a list of render settings) would take on this PC, and how long the queue is, WITHOUT starting anything.

Inputs: image, items, model, speed, +8 more.

exportB

Ready to post: the platform's shape, H.264 + AAC, file size under the limit. platform: shorts | reels | tiktok | facebook (10 MB) | youtube | x | square. max_mb overrides the size limit.

Inputs: video (required), max_mb, platform.

extend_clipB

GPU. Makes a video longer without it falling apart: the last frame becomes the start of a new clip (same model and settings, your prompt for what happens next), then both are joined. Takes a render's time.

Inputs: video (required), prompt (required), seconds, transition.

extract_audioA

Saves the video's sound as an .m4a file.

Inputs: video (required).

fadeA

Fades from black at the start and to black at the end (seconds; 0 = none). Sound fades too.

Inputs: video (required), fade_in, fade_out.

free_gpu_memoryA

Unloads models from the graphics card. Use after an out-of-memory error or before another program needs the GPU. Refused while a video renders.

Inputs: no named parameters.

get_defaultsA
Read-only

Your saved defaults for render (e.g. model, quality, style). They apply when render leaves a setting out.

Inputs: no named parameters.

get_logsA
Read-only

The last lines of the ComfyUI log (max 500). contains= keeps only lines with that text, e.g. "error".

Inputs: lines, contains.

install_custom_nodeA

Installs a ComfyUI node pack: a GitHub link, a pack name, or a node type from check_workflow. Node packs are programs from other people, so ASK THE USER FIRST, then call again with user_confirmed=true. ComfyUI needs a restart after (restart_comfyui).

Inputs: source (required), user_confirmed.

job_statusA
Read-only

Waits until the job is done (up to wait_seconds, max 55) and reports progress, then returns its status. Call again while status is queued or rendering. No id = the last 10 jobs. Also takes a batch id ("batch-...", from render_batch), a fix id ("fix-...", from auto_fix) or an edit id ("edit-...").

Inputs: id, wait_seconds.

join_clipsB

Clips into one video, in order. transition: cut | fade (through black) | crossfade; seconds = its length.

Inputs: videos (required), seconds, transition.

lipsyncB

Moves a real human face's lips to a voice (LatentSync). Needs a graphics card with 8 GB or more.

Inputs: audio (required), video (required).

list_modelsA
Read-only

What models are installed and what can be downloaded, with size, the card they need, and if we tested them. "best" marks the best one for this PC's graphics card. Use a model's "use_as" value in render(model=...).

Inputs: no named parameters.

list_videosB
Read-only

The finished videos, newest first, with their file path and short settings.

Inputs: limit.

list_workflowsB
Read-only

Your own ComfyUI workflows (API-format .json files in the workflows folder) and the inputs each one lets you change (prompt, image, seed, steps... found automatically, or nodes titled "$name" in ComfyUI).

Inputs: no named parameters.

loop_clipC

Repeats the clip (2-20 times).

Inputs: video (required), times.

make_gifD

An animated GIF of the video.

Inputs: video (required), fps, width.

make_voiceA

A voice line from text, as an audio file for add_voice. engine "edge" (default): free online voices, fast, with word timings for auto captions; voice: andrew, emma, ava, brian, es, es-f, pt, pt-f, or any edge voice name. engine "local": Chatterbox on this PC, offline, CLONES the voice in sample (a 5-15 s recording); slow.

Inputs: text (required), voice, engine, sample, +1 more.

queueA

The render queue. action: list (default) | pause (nothing new starts; the current render finishes) | resume | clear (cancels waiting jobs; priority="batch" clears only batch jobs) | front (move waiting job id first) | priority (set job id to priority "user" or "batch").

Inputs: id, action, priority.

reformatA

Changes the shape: 9:16 (Shorts, Reels, TikTok) | 1:1 | 4:5 | 16:9. mode: blur (blurred copy behind, nothing is cut off) | crop (fills the frame, cuts the sides) | pad (black bars).

Inputs: video (required), mode, shape.

remove_audioC

The same video without sound.

Inputs: video (required).

remove_modelA
Destructive

Removes a model the user added (from list_models "added_by_user"). Only our link is removed, never the user's file. Ask the user first, then call with user_confirmed=true.

Inputs: file (required), user_confirmed.

renderA

Makes a video from a prompt, starting from a picture (image = a name from upload_image) or, with no image, from the prompt alone (text to video: describe the whole scene). Everything else is optional: model: small (default, fast) | ltx13 | vace | an added model's file name (list_models) style: e.g. "Balanced", "Soft and calm", "Lively", "Slow motion" (render_options lists all) quality: "Draft" (fastest) | "Normal" (default) | "High" | "Best" (1080p, very slow) | "Manual" (then width, height) seconds: 1-7 (2-3 holds together best). camera: still | zoom | zoomout | panr | panl. speed: slow | normal | fast movement: 10-70, how much things move (style sets it). seed: -1 = new each time; same seed + settings = same video. avoid: what must not appear (vace and LTX dev models only). priority: "user" (default, goes first) or "batch" (waits while the user's own renders go first). Returns a job id right away; then call job_status(id).

Inputs: prompt (required), seed, avoid, image, +10 more.

render_batchA

Many clips at once. items = a list of render settings, e.g. [{"image": "cat.png", "prompt": "...", "style": "Lively"}, ...] (same keys as render). All items are checked before any starts. Batches wait while the user's own renders go first. Over 30 minutes in total: tell the user the estimate and call again with user_confirmed=true. Returns a batch id; job_status(batch id) reports progress.

Inputs: items (required), label, priority, user_confirmed.

render_optionsA
Read-only

The models, styles, qualities, cameras and speeds render() accepts, with what each one does.

Inputs: no named parameters.

render_variationsA

Several versions of an earlier video to pick from (2-6). vary: "seed" (same settings, different luck: the usual choice) or "movement" (less and more movement around the original, same seed). Returns job ids for job_status.

Inputs: video (required), vary, count, priority.

rerenderB

Makes a video again with the settings of an earlier one (a name from list_videos). seed: leave out for the same seed (same video, e.g. after a crash), -1 for a new variation, or a number.

Inputs: video (required), seed, prompt.

restart_comfyuiA
Destructive

Stops and starts ComfyUI again. Fixes a stuck ComfyUI. Refused while a video renders. Ask the user first, then call with user_confirmed=true.

Inputs: user_confirmed.

reverse_clipC

Plays the clip backwards.

Inputs: video (required).

review_videoA

Shows you the clip: frames spread over it (start, middle, end for 3) side by side in ONE picture, after the start picture, so you can SEE if it worked. judge=true also asks a vision model for a score and problems (uses the Gemini/NVIDIA key from OpenDaemon). Look for: the character changing, the action not happening, extra hands, melting, the end falling apart.

Inputs: video (required), judge, frames.

run_workflowA

Runs one of your workflows. inputs = {"prompt": "...", "image": "C:/pics/cat.png", "steps": 20} using the names from list_workflows; a file path for image/video/audio is copied into ComfyUI for you. A new random seed each run unless you pass seed or keep_seed=true. Checks the workflow first and refuses if it would fail (force=true runs anyway). Returns a job id: wait with job_status; the saved files come back in "files".

Inputs: name (required), force, inputs, priority, +1 more.

save_workflowB

Keeps a workflow (ComfyUI API format) under a name, so run_workflow can use it. Returns its inputs.

Inputs: name (required), workflow (required), overwrite.

search_modelsA

Searches Hugging Face for video models that fit this PC's graphics card (bigger ones are hidden). These are NOT tested by us; check_model reads a file's header first. Empty query = LTX-Video and Wan VACE models.

Inputs: query.

set_defaultsA

Saves render defaults so later renders can leave them out. clear=true forgets all of them.

Inputs: avoid, clear, model, style, +3 more.

smooth_motionB

Smoother motion: makes in-between frames, 24 -> 30/48/50/60 fps (on the processor, about 2 s per second of video).

Inputs: video (required), fps.

start_comfyuiA

Starts ComfyUI if it is not running (takes 20-60 s). Does nothing if it already runs.

Inputs: no named parameters.

suggest_settingsA
Read-only

Settings to start with for a goal ("calm portrait", "lively dance", "quick test"), plus the best-scored videos made on this PC so far (scores come from review_video(judge=true) and auto_fix).

Inputs: goal, model.

thumbnailA

A still picture (JPG) from the video at at seconds (default: the middle), for a post's cover.

Inputs: video (required), at.

title_cardB

A card at the START: text and/or a picture on a dark background, faded into the video.

Inputs: video (required), text, image, seconds.

trim_clipB

Keeps only start..end seconds. Good for cutting the end of a clip where it starts to melt.

Inputs: video (required), end, start.

upload_imageA

Puts a picture into ComfyUI's input folder. Give ONE of: path (a file on this PC), url (http/https), base64_data (the image bytes; give a name like "cat.png"). Returns the name to use in render(image=...).

Inputs: url, name, path, base64_data.

upscale_clipA

Bigger video. method "simple": resize to height with sharpening (seconds, any video). method "ltx": GPU, renders the same clip again with the LTX second pass to 1080p - real detail, but 30-50 min on a 4 GB card.

Inputs: video (required), height, method.

video_infoB
Read-only

Everything about one video: all its settings, size, length, path and a link to watch it.

Inputs: video (required).

write_avoid_listB

Writes the avoid list (negative prompt) for a prompt. Only vace and LTX dev models use it.

Inputs: image (required), prompt (required).

write_promptB

Turns a short idea ("the cat waves hello") into a full video prompt, looking at the uploaded picture. Uses the AI keys set up in OpenDaemon (Gemini or NVIDIA). You can also write the prompt yourself.

Inputs: idea (required), image (required).

Tool Schema Changelog

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

Loading tool schema changelog…

    TDQS

    B3.2/5.0

    Scored across 67 tools

    Disambiguation4/5

    Tools are grouped into clear domains (render, queue/system, model management, editing, workflows, audio) and most have distinct, well-explained purposes. The main overlaps are intentional convenience layers: render vs rerender vs render_batch vs render_variations, and build_video vs the individual join_clips/add_voice/add_captions/export tools, which could make an agent hesitate about which path to take. add_text vs title_card vs add_captions vs add_logo are also similar enough to require reading descriptions carefully.

    Naming Consistency4/5

    The dominant pattern is verb_noun (render_batch, delete_video, list_models, add_voice, check_workflow), applied consistently across most of the surface. Some noun-only outliers exist (queue, estimate, render, thumbnail, boomerang, export, fade, title_card, job_status), but they are readable and each denotes a single unambiguous operation.

    Tool Count3/5

    67 tools is heavy for a single server, even one covering a full video pipeline from model download to platform export. Several tools could be folded into aggregates that already exist (build_video already combines join/caption/voice/export), so the set feels broader than strictly necessary. The breadth is somewhat justified by the domain, but it is at the upper edge of what an agent can navigate comfortably.

    Completeness4/5

    Coverage is impressive: model install/check, rendering, batching, queue control, diagnostics, editing, audio/voice, captions, workflows, and platform export are all present with no obvious dead ends for the core workflow. Minor gaps remain, such as deleting or renaming saved workflows, and job history beyond the last 10 entries.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers