pixmax-mcp
Click on "Deploy 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., "@pixmax-mcp Generate a neon cityscape with Seedream 5.0"
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.
pixmax-mcp

A Model Context Protocol server for the Pixmax generation API. Gives any MCP client — Claude Desktop, Cursor, Claude Code, your own agent — one set of tools to generate images, video, text, audio, and 3D across dozens of models on a single key: Seedream 5, Nano Banana, GPT Image, Qwen, Kling, Veo 3.1, Hailuo, Wan, Hunyuan 3D, ElevenLabs, and more.
Self-contained and dependency-light — it talks straight to console.pixmax.ai/openapi. No other services involved.
Unofficial, community-built. Not affiliated with Pixmax.
Get a Pixmax API key
Create an account at pixmax.ai.
Click your account icon (top right) → Pixmax API → create a key. It's only shown once, so copy and save it right away.
Click Top Up (top right of the main screen) to add credits — a new account has none, and tasks will fail without balance.
Related MCP server: universal-image-mcp
Install
Requires Node 18+ and a Pixmax platform key (pk_live_…, created in the Pixmax console).
npx pixmax-mcp # run directly
# or
git clone https://github.com/huikku/pixmax-mcp && cd pixmax-mcp && npm installConfigure your MCP client
Claude Desktop
Add to claude_desktop_config.json (Settings → Developer → Edit Config):
{
"mcpServers": {
"pixmax": {
"command": "npx",
"args": ["-y", "pixmax-mcp"],
"env": { "PIXMAX_API_KEY": "pk_live_your_key_here" }
}
}
}Cursor / Claude Code / other stdio clients
Same idea — run npx -y pixmax-mcp with PIXMAX_API_KEY in the environment. For Claude Code:
claude mcp add pixmax -e PIXMAX_API_KEY=pk_live_your_key_here -- npx -y pixmax-mcpEnvironment
Variable | Required | Default | Purpose |
| ✅ | — | Your |
|
| Credit→USD rate, display only (varies by plan) | |
| auto | Reuse a specific Pixmax project | |
|
| Override the API base |
Tools
Tool | What it does |
| List models your key can use (filter by type), with estimated credit cost. Call this first. |
| Image gen. |
| Video gen. |
| Text/LLM models. Returns the text. |
| Text → 3D model ( |
| Speech & music. |
| Storyboard panels via the |
| Poll a task started with |
| Recent task history — status, prompt, and result URL(s). Recover a result you forgot to save, or audit activity. |
Every generate tool accepts save_to — a directory to download the result into, because Pixmax result URLs live on object storage and expire. Always save anything you want to keep. All of them also accept wait: false to submit and return a task id immediately (poll with get_task) — worth it for slow jobs like video and 3D.
Examples (natural language to your agent)
"List the Pixmax video models."
"Generate an image of a neon-lit alley at night with Seedream 5.0 Pro at 4K, save it to
./out."
"Using
./cleo.pngas a character reference, generate a shot of her on a rooftop in the rain with Nano Banana Pro."
"Animate
./shot.pnginto a 5-second clip with Kling V3."
Models (what's typically available)
Depends on your account. list_models is authoritative. Common ones:
Image — Seedream 5.0 Pro / Lite / 4.5 · Nano Banana / 2 / Pro · GPT Image 2 · Qwen Image Edit Plus / Max · MiniMax Image · Wan 2.7 Image / Pro
Video — Kling V3 / V3 Omni / O1 / 2.6 · Veo 3.1 · Hailuo 02 / 2.3 · Wan 2.6 / 2.7 · PixVerse C1 / V6 · Vidu Q2 / Q3 · Seedance 1.5 / 2.0
Text — DeepSeek V4 Flash / Pro · Gemini 2.5 / 3 / 3.1 · MiniMax M3 · Doubao Seed 2.x
Audio — ElevenLabs V2 / V3 / Music · MiniMax Speech / Music
3D — Hunyuan 3D Pro 3.0 / 3.1 (text-to-3D)
Model quirks the server handles for you
Veo 3.1 runs at 8s (rejects other durations).
Hailuo runs at 6s or 10s only.
Hunyuan 3D is text-to-3D only.
MiniMax Music requires
lyrics.Wan 2.7 Image models require at least one input image.
Cost
Pixmax bills in credits. Each tool reports the task's actual credit cost (from the API) and an approximate USD figure using PIXMAX_CREDIT_USD. The USD number is display-only — your real rate depends on your subscription tier. Failed and cancelled tasks cost nothing.
Pricing vs fal.ai
Many of these models are also on fal.ai, so here's how they compare. Pixmax figures are the measured credit cost at the Pro tier ($1 ≈ 143 credits, i.e. $0.007/credit); fal figures are USD list price. Both were measured mid-2026 and both providers change pricing over time — treat this as directional, not a live quote.
One structural difference matters: on Pixmax, resolution and reference images are free; on fal you pay more for higher resolution. So Pixmax's advantage widens at 2K/4K and for reference-heavy (character-consistency) work.
Images — per image
Model | Pixmax | fal.ai | |
Seedream 5.0 Lite | $0.035 (up to 5504px) | $0.035 (capped at 3072px) | same price, larger output |
Seedream 5.0 Pro | $0.07 | $0.135 @2K | ~48% cheaper |
Nano Banana | $0.028 | $0.039 | ~28% cheaper |
Nano Banana 2 | $0.098 | $0.08 @1K · $0.12 @2K | ~even at 2K |
Nano Banana Pro | $0.126 | $0.15 @2K · $0.30 @4K | −16%, −58% at 4K |
GPT Image 2 | $0.042 | — |
Video — per second
Model | Pixmax | fal.ai | |
Kling V3 | $0.077/s | $0.224/s | −66% |
Kling 2.6 | $0.028/s | $0.07/s | −60% |
Kling O1 | $0.056/s | $0.095/s | −41% |
Hailuo 2.3 | $0.049/s | $0.12/s | −59% |
Veo 3.1 | $0.175/s | $0.20/s | −13% |
Seedance 2.0 Mini | $0.098/s | $0.151/s | −35% |
Seedance 2.0 Fast | $0.154/s | $0.242/s | −36% |
Seedance 2.0 | $0.196/s | $0.302/s | −35% |
At a lower Pixmax subscription tier (Starter, $0.010/credit) the image gaps narrow; at annual tiers ($0.005/credit) they widen further. Run list_models for the live credit estimate against your own key.
Develop
npm install
PIXMAX_API_KEY=pk_live_... npm run smoke # spawns the server, lists tools, runs one real generation
npm run inspect # open the MCP InspectorLicense
MIT © John Huikku / Alienrobot LLC
Available Tools
10 toolsgenerate_3dGenerate a 3D modelA
Generate a 3D model (.glb) from a text prompt with Hunyuan 3D. Text-to-3D only (image-to-3D is not supported).
Inputs: prompt (required), wait, model, save_to, +4 more.
generate_audioGenerate audioA
Generate speech or music (ElevenLabs V2/V3/Music, MiniMax Speech/Music). For MiniMax Music pass lyrics. Cost scales with duration.
Inputs: model (required), prompt (required), wait, lyrics, +3 more.
generate_imageGenerate an imageA
Generate an image with a Pixmax model (Seedream 5 Pro/Lite, Nano Banana / 2 / Pro, Midjourney V7/V8.1/Niji, GPT Image 2, Qwen, MiniMax, Wan). Pass reference_images (file paths or URLs) for image-to-image or character consistency — up to 14, and they are free. Rich refs ({source, role, strength, note}) get their guidance woven into the prompt. style keeps a whole set visually coherent. Failed generations retry once automatically. Blocks until ready and returns the URL(s).
Inputs: model (required), prompt (required), wait, count, +7 more.
generate_storyboardGenerate a storyboard panelA
Generate an image via the GENERATE_STORYBOARD node type (GPT Image 2, Nano Banana 2, Nano Banana Pro). Identical to generate_image — same pricing, same single-image output — just classified under a different node type. No multi-panel behavior; use generate_image unless something downstream specifically requires the GENERATE_STORYBOARD classification. Pass reference_images for consistency.
Inputs: prompt (required), wait, count, model, +5 more.
generate_textGenerate textB
Run a text/LLM model (DeepSeek V4, Gemini 3.1, MiniMax, Doubao Seed). Returns the generated text.
Inputs: model (required), prompt (required).
generate_videoGenerate a videoA
Generate a video with a Pixmax model (Kling V3/O1/2.6, Veo 3.1, Hailuo, Wan, PixVerse, Vidu, Seedance). Pass an image (path or URL) for image-to-video, plus last_frame for start+end keyframe control (the model interpolates between them — Seedance 2 and Hailuo). Seedance 2 (PIXDANCE_2 / _FAST / _MINI) additionally supports multimodal references — reference_videos (borrow motion/style from a clip) and reference_audio (drive speech/lip-sync); audio must be paired with an image or video reference. Other models take images only. Note: Veo 3.1 runs 8s; Hailuo runs 6s or 10s. Video can take a few minutes; set wait=false to get a task id and poll with get_task.
Inputs: model (required), prompt (required), wait, count, +12 more.
get_taskGet task status / resultA
Check a task started with wait=false. Returns status, and result URL(s) + cost once COMPLETE.
Inputs: task_id (required), save_to.
list_modelsList Pixmax modelsA
List the models your Pixmax key can use, with type, estimated credit cost, and reference capacity. Call this first to discover exact model codes. Pass model to get one model's full prompting + reference guidance (best-for, prompt style, ref limits) before generating. Optionally filter by node type.
Inputs: model, node_type.
list_tasksList recent tasksA
List your recent Pixmax generation tasks (newest first) with status, model, prompt, and result URL(s). Useful for recovering a result you forgot to save, checking on wait=false jobs, or auditing recent activity.
Inputs: limit, status.
prompting_guidePrompting guideA
Get the prompting craft rules for a medium (video, image, audio, 3d) — how to structure a prompt, camera/motion vocabulary for video, how to use references, and the common failure modes. Pass a model to also get that model's specific tips and hard limits. Call this BEFORE writing a prompt for an unfamiliar medium or model — video especially has rules that silently ruin output (e.g. implying a cut makes the model morph instead).
Inputs: model, category.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 10 tools
Most tools have distinct purposes, but generate_image and generate_storyboard are described as nearly identical, differing only in node type. This overlap could cause confusion for an agent.
Tool names follow a consistent verb_noun pattern (e.g., generate_image, list_models, get_task). The only deviation is 'prompting_guide', which uses a gerund instead of a verb, but it is still clear.
With 10 tools covering a broad range of generative model capabilities (image, video, text, 3D, audio) plus management functions, the count is well-scoped and not overwhelming.
The tool surface covers generation and monitoring for all advertised media types. Minor gaps include the lack of a cancel task or delete result tool, but core workflows are complete.
Related MCP Connectors
MCP server for Pixapi: check live credit pricing and balance, then generate images and video.
MCP server for Google Veo AI video generation
MCP server for Wan AI video generation
MCP server for Qwen Image 3 AI image generation
Related MCP Servers
- AlicenseAqualityDmaintenanceMCP server for Krea.ai enabling AI image and video generation using models like Flux, Hailuo, Runway, Kling, and others via natural language prompts.933 npm11Apache 2.0
- AlicenseAqualityDmaintenanceMCP server for multi-provider AI image generation (AWS Bedrock, OpenAI, Google Gemini) enabling image generation, transformation, and editing through a unified interface.41MIT
- AlicenseNot gradedqualityDmaintenanceMCP server for MiniMax's multimodal generation models, enabling text-to-speech, voice cloning, image, video, and music creation through natural language.MIT
- AlicenseAqualityDmaintenanceMCP server for AI-powered media generation: images, videos, audio, and upscaling using 99 AI models.6MIT