A
licenseA
qualityNot graded
maintenanceEnables interaction with MiniMax AI APIs for text-to-speech, voice cloning, video generation, image generation, and music creation through MCP clients like Claude Desktop and Cursor.
9
comfy-h3-mcp
by jonpojonpo
This server drives MiniMax-H3 video and audio generation on a local ComfyUI instance. You can:
Generate videos using
h3_image_to_video(text-to-video or keyframe-guided with first/last frame) andh3_reference_to_video(prompt plus reference images, videos, or audio for consistency).Manage jobs with
job_status(poll progress and get output URLs), andjob_cancel(cancel pending or running jobs).Preview results via
job_preview(contact sheet of frames and soundtrack).List assets with
list_assets(available models, input files, and system settings like sage attention).
Provides a tool to generate a keyframe image using OpenAI's GPT Image API, which can be used as the first frame for H3 video generation.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@comfy-h3-mcpGenerate a video from the prompt: a dragon soaring over a castle"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
comfy-h3-mcp
A small MCP server for driving MiniMax-H3 video+audio generation on a local ComfyUI. Eight tools, not a general ComfyUI control plane — the point is to keep the agent's context cost near zero for the one thing this rig actually does.
Tools
Tool | What it does |
| Text-to-video, or keyframe-guided via |
| Prompt + reference images / videos / audio (ref2va model) |
| Poll a |
| Drop from queue if pending, interrupt if running |
| Return a contact sheet and soundtrack for a finished clip |
| Save a frame of a finished clip as a reusable reference still |
| Generate and upload a GPT Image keyframe (metered API) |
| H3 models present, and input files usable by name |
Generation takes minutes, so the two generate tools submit and return a
prompt_id immediately. There is no session state — the prompt_id is the
only handle, and ComfyUI already owns it.
Related MCP server: ComfyUI MCP Server
Acceleration presets
Both generation tools accept one acceleration value:
Preset | Stack | Use |
| Native H3 path (global Sage still applies) | Native-trajectory A/B reference |
| Sol-Attn, | Quality-sensitive accelerated renders |
| Sol-Attn, | Normal local generation |
| Sol-Attn, | Prompt and motion drafts |
These are inference-time approximations. They do not retrain or modify the
checkpoint, but they can change motion, fine detail, and audio for the same
seed. Keep off available as the reference and re-run important shots with
quality or off when fidelity matters.
The Sol-Attn custom node must be installed in ComfyUI:
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-SolAttn_triton.gitlist_assets reports whether both SolAttnPatch and native EasyCache are
loaded. Sol-Attn's Triton kernels compile for each new tensor shape, so the
first run at a resolution/duration can be slower than subsequent warm runs.
ComfyUI GUI workflow
Load workflows/minimax_h3_sol_balanced.json
from ComfyUI's Workflows → Open menu. It is derived from ComfyUI's official
MiniMax-H3 text/image-to-video template and inserts the measured balanced path:
UNETLoader → SolAttnPatch → EasyCache → scheduler + guiderThe workflow opens at 0.4 MP, 5 seconds, and 20 RES multistep steps. Global
SageAttention remains controlled by the ComfyUI launch flag. The note inside
the workflow gives the quality, balanced, and fast settings.
Setup
uv venv && uv pip install -e .
claude mcp add comfy-h3 -s user \
-e COMFYUI_URL=http://127.0.0.1:8188 \
-- /path/to/comfy-h3-mcp/.venv/bin/comfy-h3-mcpCOMFYUI_URL defaults to http://127.0.0.1:8188.
The optional gpt_image_first_frame tool also needs OPENAI_API_KEY in the
environment that launches the MCP server. It uses the metered GPT Image API;
the local H3 generation tools do not require that key.
Timing
Generation takes minutes. Every submit returns estimated_seconds plus a
suggested poll interval, so a client knows the difference between "slow" and
"stuck". Measured on an RTX 4090 at 124 frames:
Config | Time |
864×480, 20 steps, sage | 3m 43s |
864×480, 20 steps, sage + | 2m 33s |
864×480, 20 steps, sage + | 1m 52s |
864×480, 24 steps, sage | 4m 11s |
1344×768, 30 steps, sage | 13m 13s |
1344×768, 30 steps, no sage | 18m 02s |
The estimator scales with pixels × steps × length. Its accelerated factors are
fitted to the two local measurements above. balanced skipped 5/20 denoiser
evaluations; fast skipped 7/20. Roughly 30 s is fixed model load/decode work,
which these DiT optimizations do not remove — the nvfp4 text-encoder path is
emulated on this hardware.
Draft small, finish large. 864×480 is the template's draft setting; the model's documented full-quality 16:9 target is ~1.0 MP (1344×768). Iterate prompts at the default, then re-run keepers at 1344×768 with the same seed — noting that a seed does not guarantee an identical image across a resolution change, only a related composition.
Sage attention
list_assets reports sage_attention.global, detected by inspecting the ComfyUI
process for --use-sage-attention (local servers only; None when it can't be
determined). When sage is global, the per-call sage_attention parameter is
redundant — the tool descriptions tell clients not to set it and not to
recommend enabling sage, which otherwise happens: an agent that can't see the
launch flag will report a normal run as "grinding" and advise a no-op fix.
Sage gives roughly 1.36× here. It is a lossy approximation: same seed produces a different sample, not the same one faster (SSIM 0.78 on luma, audio differs too). Keep it in a fixed state once you lock a seed.
Required models
Loaded by name, so they must be present:
models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensorsmodels/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensorsmodels/vae/minimax_h3_video_vae_fp16.safetensorsmodels/vae/minimax_h3_audio_vae_fp32.safetensorsmodels/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
list_assets reports anything missing.
Notes on the model's constraints
These are enforced in graphs.py, mirroring comfy_extras/nodes_minimax_h3.py,
so callers see the real numbers up front rather than discovering that ComfyUI
snapped them:
Length is a frame count at 24 fps, snapped up to the
17k+5grid. 124 ≈ 5.2 s. Trained range is roughly 124–362; longer is untested.Canvas is capped at 768×1344 pixels of area with each axis rounded to 32. Oversized requests are re-fitted via the model's own
adapt_canvasrule, so 1920×1080 becomes 1344×768.References are addressed positionally in the prompt as
<Picture i>,<Video k>,<Audio j>— all 1-based per type. Max 9 images, 3 videos, 3 audio. A reference video's soundtrack is wired through automatically.Prefer stills to reference videos.
ref_videoscarry identity poorly and drag their own soundtrack into the output, fighting the audio the prompt asked for. Usegrab_referenceto lift a frame from an earlier clip instead: runjob_preview, pick a tile off the contact sheet, andgrab_reference( prompt_id, tile=N)saves that exact source frame at full resolution into ComfyUI's input folder, ready to pass asref_images.ref_image_size="max"uses a 2048 px short edge for better identity fidelity but is several times slower, because reference tokens ride through every sampling step.
Defaults
These mirror the official video_minimax_h3_t2v.json template, verified
node-for-node:
Value | |
Resolution | 864×480 — 16:9 at 0.4 MP, rounded to 32 |
Steps | 20 |
Sampler |
|
Scheduler |
|
Guidance |
|
Sigma shift | none — |
Pass width/height explicitly to override the megapixel calculation, or
raise megapixels. Note the cost: 1344×768 is 2.5× the pixels of the default,
and res_multistep is a higher-order sampler, so 20 steps here is not a
downgrade from 30 steps of euler — it is faster and comparable in quality.
shift_video/shift_audio default to None, which omits MiniMaxH3SigmaShift
entirely. Set either one to insert the node and override the model default.
Graph shape
UNETLoader ─► Sol-Attn ─► EasyCache ─┐
CLIPLoader(minimax) ─┐ ├─► BasicGuider ─┐
VAELoader(video) ────┼─► MiniMaxH3{ImageToVideo, │
VAELoader(audio) ────┘ ReferenceToVideo} │
│ positive ───────────────┘
└─ latent ────────────────┐
│
RandomNoise ─┐ │
KSamplerSelect(res_multistep) ─┼─► SamplerCustomAdvanced ◄┘
BasicScheduler(simple, 20) ────┘ │
├─► VAEDecode(video vae) ─────┐
└─► VAEDecodeAudio(audio vae) ┤
▼
CreateVideo(24fps) ─► SaveVideoBoth VAEs read the joint AV latent directly — the nested video/audio pair
needs no explicit split node. (LTXVSeparateAVLatent does work here despite the
name, but the template doesn't use it and neither do we.)
Autogrow reference inputs serialize as dotted API keys
(ref_images.ref_image_0, ref_video_audios.ref_video_audio_0), per
finalize_prefix() in comfy_api/latest/_io.py.
Requires
MCP SDK 2.0+ (MCPServer; FastMCP was removed).
Available Tools
6 toolsh3_image_to_videoA
Generate video with synchronized audio from a text prompt, optionally anchored by a first and/or last keyframe (MiniMax-H3 fl2va model).
With no keyframes this is pure text-to-video. Images may be a local file path (uploaded automatically) or a name already in ComfyUI's input folder. length is a frame count at 24 fps and snaps up to the model's 17k+5 grid; 124 frames is about 5 seconds, and the trained range is roughly 124-362. Leave width/height unset: they default to 864x480, which is both faster and closer to the prompt than larger canvases. Only raise it if the user asks. This produces ONE continuous shot - it cannot contain cuts, so for a multi-shot sequence call this once per shot and edit the clips together rather than describing several shots in one prompt. Takes MINUTES: ~3-4 min at the defaults on an RTX 4090, longer at higher resolution or step count. The response carries estimated_seconds; poll job_status at that cadence instead of assuming a long run has hung. Do NOT set sage_attention without first checking list_assets - if ComfyUI runs with --use-sage-attention, sage is already on and this is redundant. Returns immediately with a prompt_id - poll job_status to get the output.
ParametersJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| seed | No | ||
| steps | No | ||
| width | No | ||
| height | No | ||
| length | No | ||
| prompt | Yes | ||
| scheduler | No | simple | |
| last_frame | No | ||
| megapixels | No | ||
| first_frame | No | ||
| shift_audio | No | ||
| shift_video | No | ||
| sampler_name | No | res_multistep | |
| sage_attention | No | disabled | |
| filename_prefix | No | video/h3_i2v |
TDQS
A4.4/5.0
Behavior5/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description fully discloses operational behavior: execution time ('~3-4 min at the defaults on an RTX 4090'), immediate return with a prompt_id, polling recommendation using estimated_seconds, the single-continuous-shot limitation, resolution defaults and trade-offs, frame length constraints, and the sage_attention redundancy caveat. This is exceptional transparency for a complex tool.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness5/5
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is longer than typical but every sentence earns its place. It is front-loaded with the core purpose, then delivers practical operational details, performance expectations, and caveats in a structured, scannable format. No redundancy or filler.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness5/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 15-parameter tool with no output schema, this description is remarkably complete. It covers the execution flow (returns prompt_id, poll job_status), performance, constraints, parameter defaults, and gotchas. The only minor omission is detailed error handling, but the provided context is sufficient for an agent to invoke and monitor the tool correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters4/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, so the description compensates by explaining the most significant parameters: length (frame count, 24fps, grid snapping, range), width/height (defaults and advice), sage_attention (do not set without list_assets), and first_frame/last_frame (local path or input-folder name). However, several parameters like steps, seed, scheduler, sampler_name, megapixels, shift_audio/video, and filename_prefix are left unexplained, though many are conventional.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose4/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states a specific action: 'Generate video with synchronized audio from a text prompt, optionally anchored by a first and/or last keyframe (MiniMax-H3 fl2va model)'. It distinguishes between text-to-video and image-anchored modes but does not explicitly differentiate from the sibling tool h3_reference_to_video, which is a similar generation tool.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines4/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
Provides concrete usage context such as 'With no keyframes this is pure text-to-video' and warns against multi-shot use: 'cannot contain cuts, so for a multi-shot sequence call this once per shot and edit the clips together'. Also advises checking list_assets before setting sage_attention. Does not explicitly name alternative tools, but clear situational guidance is present.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
h3_reference_to_videoA
Generate video with synchronized audio from a prompt plus reference images, videos, and/or audio (MiniMax-H3 ref2va model).
References are addressed positionally in the prompt as , and , all 1-based per type - e.g. "<Picture 1> walks through the door speaking in the voice of <Audio 1>". Max 9 images, 3 videos, 3 audio. A reference video's own soundtrack is passed through automatically. ref_image_size "match" scales references to the output's pixel area; "max" uses a 2048px short edge for better identity fidelity but is several times slower, since reference tokens ride through every step. This is the tool for CONSISTENCY ACROSS SHOTS: generate shot 1, pull a frame of the character or location from it, and pass it here as a reference for shots 2..n so they match. Leave width/height unset (864x480 default). Takes MINUTES; see estimated_seconds in the response and poll at that cadence. Check list_assets before setting sage_attention - it is redundant when ComfyUI already runs with --use-sage-attention. Returns immediately with a prompt_id - poll job_status to get the output.
ParametersJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| seed | No | ||
| steps | No | ||
| width | No | ||
| height | No | ||
| length | No | ||
| prompt | Yes | ||
| scheduler | No | simple | |
| megapixels | No | ||
| ref_audios | No | ||
| ref_images | No | ||
| ref_videos | No | ||
| shift_audio | No | ||
| shift_video | No | ||
| sampler_name | No | res_multistep | |
| ref_image_size | No | match | |
| sage_attention | No | disabled | |
| filename_prefix | No | video/h3_ref2v |
TDQS
A4.3/5.0
Behavior4/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description discloses key behaviors: the job is asynchronous ('Returns immediately with a prompt_id - poll job_status'), it takes minutes with an estimated_seconds field, and ref_image_size 'max' is 'several times slower.' It also warns about sage_attention redundancy. It doesn't cover permissions or failure modes, but the disclosure is strong for a generation tool.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness4/5
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a dense single paragraph where every sentence adds value: purpose, syntax, limits, parameter trade-offs, use case, timing, and async behavior. While long, it is justified by complexity. Structure could be improved with bullets, but it is not wasteful.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness4/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 17-parameter tool with no annotations and no output schema, the description covers the critical invocation details: async flow, reference syntax, parameter trade-offs, and the cross-shot workflow. It doesn't explicitly say references are asset IDs from list_assets, but the sibling context implies it. Overall, sufficient for correct use.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters4/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Despite 0% schema coverage, the description explains the most important parameters: positional refs in prompt, max counts for ref arrays, ref_image_size trade-offs, width/height defaults, and sage_attention. It omits steps, scheduler, megapixels, and shift_* but these are secondary with defaults. The compensation is solid but not exhaustive.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose5/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific verb and resource: 'Generate video with synchronized audio from a prompt plus reference images, videos, and/or audio.' It also explicitly identifies this as the tool for 'CONSISTENCY ACROSS SHOTS,' which distinguishes it from the sibling h3_image_to_video. The positional reference syntax and model name add further specificity.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines4/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides a concrete use case workflow: generate shot 1, pull a frame, and pass it as a reference for subsequent shots. It also gives guidance on leaving width/height unset and checking list_assets before setting sage_attention. It does not explicitly contrast with h3_image_to_video, but the cross-shot scenario clearly implies when this tool is appropriate.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
job_cancelA
Cancel an H3 job: removes it if still queued, interrupts it if running.
ParametersJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| prompt_id | Yes |
TDQS
A3.8/5.0
Behavior4/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It transparently explains the two possible outcomes (removal if queued, interruption if running), which is valuable. It does not cover edge cases like already-completed jobs or error handling, but for a cancel action, this level of detail is reasonable.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness5/5
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a concise two-clause sentence: 'Cancel an H3 job: removes it if still queued, interrupts it if running.' Every word adds value, with no redundant information or filler.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness3/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
The tool is simple with one parameter and no output schema. The description explains the core behavior but omits details about return values, error scenarios, or job states like 'completed'. While not critical for a cancel tool, the absence of these details leaves the agent with incomplete expectations.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters2/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The description does not explain the 'prompt_id' parameter beyond its name, and the schema provides no descriptions (0% coverage). The parameter name is somewhat self-explanatory, but the description fails to clarify its format, origin, or how to obtain it. This leaves a gap in agent understanding.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose5/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: 'Cancel an H3 job'. It also specifies behavior based on job state ('removes it if still queued, interrupts it if running'), which distinguishes it from sibling tools like job_status and h3_image_to_video. The verb+resource structure makes the action unambiguous.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines3/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies the primary use case: cancel a job. However, it does not explicitly mention when to choose this tool over alternatives (e.g., job_status to check status first) or any exclusions (e.g., cannot cancel completed jobs). The context is clear but lacks explicit guidance.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
job_previewA
Look at and listen to a finished H3 job.
MCP has no video content type, so this returns a contact sheet of frames sampled evenly across the clip (viewable as an image) plus the soundtrack as audio. Use it to actually judge a result before iterating on the prompt.
ParametersJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| rows | No | ||
| columns | No | ||
| prompt_id | Yes | ||
| tile_width | No | ||
| include_audio | No |
TDQS
A4.1/5.0
Behavior4/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
Discloses that MCP has no video content type, so the tool returns a contact sheet of frames plus audio instead. This is important behavioral context not available from annotations (which are absent). It implies a read-only operation and describes the return format.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness5/5
Is the description appropriately sized, front-loaded, and free of redundancy?
Two concise paragraphs with the purpose in the first sentence. The additional details about the output format and usage context are relevant and not redundant.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness4/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given no output schema, the description explains the return format (contact sheet + audio) and the reason for this design. It doesn't describe each parameter in detail, but the tool is relatively simple and the core behavior is covered.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters2/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema has no parameter descriptions (0% coverage), and the description does not explain the meaning of rows, columns, tile_width, or include_audio. Parameter names are somewhat self-explanatory, but the description adds no detail, leaving the agent to infer from names alone.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose5/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: to look at and listen to a finished H3 job. It differentiates from siblings like job_status and job_cancel by focusing on previewing results and explains the output format.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines4/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicitly says to use it to judge a result before iterating on the prompt, indicating appropriate timing. It does not name alternatives or when-not-to-use scenarios, but the context is clear enough for selection.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
job_statusA
Check a submitted H3 job. Returns queued / running / completed / failed, plus the output file paths once it has finished.
ParametersJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
| prompt_id | Yes |
TDQS
A3.8/5.0
Behavior3/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden. It discloses that the tool returns status and output file paths post-completion, and the verb 'Check' implies a read-only operation. However, it does not disclose behavior on error, missing IDs, or explicitly confirm no side effects, so it is minimally transparent.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness5/5
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is one concise sentence that front-loads the purpose ('Check a submitted H3 job') and immediately follows with the return values. Every word earns its place with no redundancy.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness4/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
The tool is simple with one parameter and no output schema. The description covers the core purpose and return values (statuses and output paths), which is sufficient for a straightforward status check. However, it omits error handling details, which prevents a perfect 5.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters3/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema has 0% coverage, and the description only mentions 'submitted H3 job' which implies prompt_id is the job identifier but does not explicitly define it. For a single parameter, this provides some additional meaning over the schema's bare 'Prompt Id' label, but it does not fully compensate for the lack of schema description.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose5/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states 'Check a submitted H3 job' with a specific verb and resource, and it distinguishes itself from siblings like job_cancel (cancel) and job_preview (preview) by focusing on status retrieval. It also lists the exact statuses returned.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines3/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description implies usage when you want the status of a submitted H3 job, but it does not explicitly state when not to use it or mention alternatives. It provides clear context but lacks exclusionary guidance, so it falls short of a 4.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
list_assetsA
List what this ComfyUI can actually load: H3 models, and the images, videos and audio already sitting in the input folder (usable by name).
ParametersJSON Schema
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
TDQS
A4.5/5.0
Behavior4/5
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
The description discloses the behavioral scope: it lists models and media files from the input folder, implying read-only access. With no annotations, this is valuable context. It doesn't explicitly state 'read-only' or return format, but the description adequately conveys the tool's behavior for a simple list operation.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Conciseness5/5
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, front-loaded sentence that immediately states the action and scope. Every word earns its place with no redundancy.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Completeness5/5
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a zero-parameter discovery tool, the description is complete: it fully explains what assets are returned and that they are usable by name. No output schema or annotations are needed, and the description provides sufficient context for an agent to decide when to invoke it.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Parameters4/5
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The tool has zero parameters, so the baseline is 4. The description compensates by clarifying what will be listed, which is the only relevant semantic context.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Purpose5/5
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb ('List') and clearly identifies the resource (loadable assets: H3 models, images, videos, audio in the input folder). It also distinguishes itself from sibling tools by focusing on asset inventory rather than job management or generation.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Usage Guidelines4/5
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description clearly establishes when to use the tool (when you need to know what ComfyUI can load). It does not explicitly mention alternatives, but sibling tools are obviously job-oriented, so context is sufficient without explicit exclusions.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
TDQS
A4.1/5.0
Disambiguation5/5
Each tool has a clearly distinct role: job management (status, cancel, preview), asset discovery, and two generation tools with different use cases (keyframes vs. reference-based consistency). The generation tools are well-separated by their descriptions, leaving no ambiguity.
Naming Consistency4/5
Names follow a predictable pattern within categories: job_* for job operations, h3_* for generation, and list_assets for assets. However, there is no single uniform verb_noun convention across all tools (e.g., job_cancel vs. list_assets), which is a minor deviation.
Tool Count5/5
Six tools is well-scoped for this focused video-generation server. Each tool fills a necessary part of the workflow—submitting generation jobs, managing jobs, inspecting results, and listing available assets—with no redundant tools.
Completeness4/5
The tool set covers the full generation lifecycle: submit (two variants), poll status, cancel, preview, and list assets. A minor gap is the lack of a listing mechanism for all submitted jobs, which could be a workaround but is not critical.
Maintenance
ActivityMaintained
ResponsivenessNo issues
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Connectors
MCP server for MiniMax H3 multimodal video generation
Create images and videos from prompts, with options for image mixing, reference images, and start/…
Generate images, video, music, voice and 3D through one API. 30 tools, 200+ models.
MCP server for Hailuo (MiniMax) AI video generation
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceDynamically loads ComfyUI workflows as MCP tools, enabling AI assistants to generate images, videos, and audio by executing workflows across categories like text-to-image, image-to-video, and text-to-audio with automatic parameter mapping and progress monitoring.141,9053MIT
- AlicenseNot gradedqualityNot gradedmaintenanceEnables AI agents to generate and iteratively refine images, audio, and video by interacting with a local ComfyUI instance through natural conversation. It provides comprehensive tools for workflow management, node introspection, and publishing generated assets.
- FlicenseBqualityDmaintenanceConnects AI assistants to ComfyUI for image, video, and audio generation, providing full control over ComfyUI through 40+ tools including quick generation, cloud API nodes, and custom workflow building.46
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/jonpojonpo/comfy-h3-mcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server