MCP Server

OpenTryOn ships a Model Context Protocol server under mcp-server/. Every model in tryon.cli.registry becomes an MCP tool automatically — the same surface as the opentryon CLI, via tryon.cli.runner.invoke_model().

Current release: OpenTryOn v0.0.5+ (pip install -U opentryon).

This page is the Docusaurus guide for the server. Keep it next to:

In-repo README (full tool table): mcp-server/README.md
TryOn Studio (the web MCP client): TryOn Studio

Why it matters

Agents in Cursor, Claude Desktop, or TryOn Studio call try-on, generate, edit, video, understand, and bg-remove tools directly.
Studio chat goes through planner_agent: it classifies intent, then runs a filtered slice of those same registry tools via invoke_model. Capability screens skip the planner and call the model tools themselves.
New registry models appear as tools with zero hand-written MCP wrappers.
CLI and MCP cannot drift — one runner, one registry.

Install

cd opentryon
pip install -e .              # core (API-backed) models
# optional GPU extras (leffa, catvton, kimi-vl, qwen3.8, ben2, …):
pip install -e ".[local]"

cd mcp-server
pip install -r requirements.txt

Copy repo-root env.template to .env and fill the keys you plan to use. The server and every adapter read that same file.

Run

# stdio — what Claude Desktop / Cursor expect
python server.py

# streamable-HTTP — required by TryOn Studio
python server.py --transport http --host 127.0.0.1 --port 8000

On startup the server prints a configuration report to stderr (which keys are set, which models are ready) — the same text opentryon_status returns at runtime.

Transport	Typical client	Endpoint
stdio (default)	Cursor, Claude Desktop	process stdin/stdout
HTTP	TryOn Studio, FastMCP `Client` over the network	`http://127.0.0.1:8000/mcp`

Studio’s only URL is OPENTRYON_MCP_URL=http://127.0.0.1:8000/mcp. Remote MCP hosts are out of scope for Studio.

Clients

TryOn Studio

HTTP MCP plus a Next.js UI (Agent, Connect, Image, VTON, Understand, Video, BG Remove). Full setup: TryOn Studio.

Cursor

Add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global). Example: mcp-server/examples/cursor_mcp_config.json.

{
  "mcpServers": {
    "opentryon": {
      "command": "python",
      "args": ["/absolute/path/to/opentryon/mcp-server/server.py"]
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json. Example: mcp-server/examples/claude_desktop_config.json.

{
  "mcpServers": {
    "opentryon": {
      "command": "python",
      "args": ["/absolute/path/to/opentryon/mcp-server/server.py"]
    }
  }
}

Python (FastMCP client)

See mcp-server/examples/example_usage.py.

import asyncio
from fastmcp import Client

async def main():
    async with Client("server.py") as client:
        result = await client.call_tool("vton_flux_vto", {
            "person": "model.jpg",
            "garment": "garment.jpg",
            "dry_run": True,
        })
        print(result.data)

asyncio.run(main())

Discovery, keys, and the planner

Always available, independent of which models you configured:

Tool	Role
`list_opentryon_tools`	Services, models, MCP tool names, env readiness
`opentryon_status`	Same human-readable report printed on startup
`list_api_keys` / `set_api_keys`	Inspect or upsert host `.env` keys (never returns secret values). Studio Connect uses these
`planner_agent`	Studio Agent chat entrypoint. Cheap LLM classifies intent, then `invoke_model` on a filtered slice. Planner Agent

Every generated model tool also accepts dry_run and output_dir, matching the CLI.

Selected tools

The tables below highlight newer families. The complete generated list lives in mcp-server/README.md and grows with tryon/cli/registry.py.

Understand tools (including Qwen3.8 and Hy4)

Multimodal image/video understanding tools include Kimi, LLaVA-NeXT, the Qwen3.8 dual path (+ Qwen3.8-Omni-Flash for audio), GLM-5.3-FlashX, Ternary Bonsai 2 27B (local server), and Hy4 preview (TokenHub LLM + local vLLM/SGLang):

MCP tool	Backend	Needs
`understand_qwen3_8_max`	DashScope Qwen3.8-Max (text/image/video, thinking + `reasoning_effort`)	`DASHSCOPE_API_KEY`
`understand_qwen3_8_omni_flash`	DashScope Qwen3.8-Omni-Flash (adds audio; text/image/audio/video, 1M context)	`DASHSCOPE_API_KEY`
`understand_qwen3_8`	Local `Qwen/Qwen3.8-27B`	`pip install opentryon[local]` + GPU
`understand_glm_5_3_flashx`	Z.ai GLM-5.3-FlashX (text/image/video, 200 tok/s)	`ZAI_API_KEY`
`understand_ternary_bonsai_2_27b`	PrismML Ternary Bonsai 2 27B (self-hosted llama.cpp/MLX server)	`BONSAI_BASE_URL` (default localhost:8080)
`understand_hy4_preview`	Tencent Hy4 preview (TokenHub LLM)	`TOKENHUB_API_KEY`
`understand_hy4_preview_local`	Hy4 via local vLLM/SGLang OpenAI server	`HY4_BASE_URL` (default localhost:8000)

Qwen-Image tools (generate / edit / VTON)

Same DashScope key as Qwen3.8-Max. Image generation is Qwen-Image, not the Qwen3.8 VLM:

MCP tool	Backend	Needs
`generate_qwen_image`	Qwen-Image 3.0 T2I (default `qwen-image-3.0-pro`)	`DASHSCOPE_API_KEY`
`edit_qwen_image`	Qwen-Image I2I (1–3 refs)	`DASHSCOPE_API_KEY`
`vton_qwen_image`	Person + garment composition	`DASHSCOPE_API_KEY`

Typical loop: understand_qwen3_8_max captions a garment, then generate_qwen_image or vton_qwen_image uses that description.

Local Diffusers twin (pip install opentryon[local] + CUDA):

MCP tool	Backend	Needs
`generate_qwen_image_local`	`Qwen/Qwen-Image-2512` T2I	GPU + recent Diffusers
`edit_qwen_image_local`	`Qwen/Qwen-Image-Edit-2511` I2I	GPU + recent Diffusers
`vton_qwen_image_local`	Edit-Plus person + garment	GPU + recent Diffusers

See Qwen-Image local.

Local dedicated VTON (Leffa + CatVTON)

MCP tool	Backend	Needs
`vton_leffa`	franciszzj/Leffa (CVPR 2025)	GPU + `opentryon[local]`
`vton_catvton`	zhengchong/CatVTON (ICLR 2025, CC BY-NC-SA)	GPU + `opentryon[local]`

See Leffa and CatVTON.

Qwen3.8 is a native multimodal / coding / agent family; OpenTryOn’s MCP tools expose the understand entry point (image and/or video + prompt). Full capability notes: Qwen3.8-Max, Qwen-Image, and Qwen3.8 local.

MiniMax H3 tools (video)

Same MINIMAX_API_KEY as Hailuo 2.3. H3 is a dual path (hosted V2 vs local Diffusers); Hailuo 2.3 stays API-only. H3 Max is the hosted fast variant (no local twin).

MCP tool	Backend	Needs
`video_generate_minimax_h3`	MiniMax H3 official V2 API (T2V / I2V / R2V, 4–15s, 768P/2K)	`MINIMAX_API_KEY`
`video_generate_minimax_h3_max`	MiniMax H3 Max (fast V2; T2V / I2V, 5–15s, 480P/768P)	`MINIMAX_API_KEY`
`video_generate_fal_h3_max`	Fal-hosted H3 Max (T2V / I2V / R2V, 5–15s, 480P/768P)	`FAL_KEY`
`video_generate_fal_h3_max_lipsync`	Fal-hosted H3 Max Lip Sync (image + audio → dubbed video, 480P–2K)	`FAL_KEY`
`video_generate_minimax_h3_local`	Open-weight `MiniMaxAI/MiniMax-H3` (768p H3-Base)	`pip install opentryon[local]` + CUDA + Diffusers from main
`video_generate_hailuo_2_3`	MiniMax Hailuo 2.3 (V1)	`MINIMAX_API_KEY`

See MiniMax H3 API, MiniMax H3 Max (Fal), and MiniMax H3 local.

NVIDIA NIM tools (understand + video)

Same NVIDIA_API_KEY for Nemotron Omni, Cosmos 3 Reasoner, and Cosmos 3 Generator.

MCP tool	Backend	Needs
`understand_nemotron_omni`	Nemotron 3 Nano Omni (image / video / audio)	`NVIDIA_API_KEY`
`understand_cosmos3_reasoner`	Cosmos 3 Reasoner (physical-world VLM)	`NVIDIA_API_KEY`
`video_generate_cosmos3`	Cosmos 3 Generator nano (T2V / I2V)	`NVIDIA_API_KEY`

See NVIDIA NIM.

Google Virtual Try-On (Vertex)

Dedicated person + product try-on. Not GEMINI_API_KEY / Nano Banana.

MCP tool	Backend	Needs
`vton_google_vton`	Vertex `virtual-try-on-001`	`GOOGLE_CLOUD_PROJECT` + ADC

See Google Virtual Try-On.

OutfitAnyone-Plus (DashScope, Beijing)

Dedicated Alibaba try-on. Not Qwen-Image composition. Needs a China Beijing-region DASHSCOPE_API_KEY.

MCP tool	Backend	Needs
`vton_outfitanyone_plus`	`aitryon-plus`	Beijing `DASHSCOPE_API_KEY`

See OutfitAnyone-Plus.

Photoroom (Virtual Try-On / Virtual Model)

Image Editing API Plus. Same PHOTOROOM_API_KEY. Prefix with sandbox_ for watermarked tests.

MCP tool	Backend	Needs
`vton_photoroom_vton`	Shopper try-on (`virtualModel.model.custom`)	`PHOTOROOM_API_KEY`
`vton_photoroom_virtual_model`	Catalog on-model (preset or custom)	`PHOTOROOM_API_KEY`

See Photoroom.

Muse Image tools (generate / edit / VTON)

First-party Meta Model API (MODEL_API_KEY). No local twin. Muse Video is not on the API yet.

MCP tool	Backend	Needs
`generate_muse_image`	Muse Image T2I (`muse-image-1.0`)	`MODEL_API_KEY`
`edit_muse_image`	Muse Image I2I / multi-ref	`MODEL_API_KEY`
`vton_muse_image`	Person + garment composition	`MODEL_API_KEY`

See Muse Image and Muse Video (not available).

P-Image-Ideogram (generate)

Same PRUNA_API_KEY as the rest of the Pruna family. Not Ideogram 4.0 (generate_ideogram / IDEOGRAM_API_KEY).

MCP tool	Backend	Needs
`generate_p_image_ideogram`	Pruna `Model: p-image-ideogram` (thinking very-low–very-high, 1K/2K)	`PRUNA_API_KEY`

See P-Image-Ideogram.

TryOn Studio — Next.js MCP client (Agent, Connect, capability screens)
mcp-server/README.md — architecture notes and full generated tool table
Planner Agent
Unified CLI
Configuration
Adding a new model
Qwen3.8-Max understanding
Hy4 preview TokenHub
Hy4 local vLLM/SGLang
Qwen-Image generate / edit / VTON
Qwen-Image local model
Qwen3.8 local model
MiniMax H3 API
MiniMax H3 Max (Fal)
MiniMax H3 local
Muse Image
P-Image-Ideogram

Why it matters​

Install​

Run​

Clients​

TryOn Studio​

Cursor​

Claude Desktop​

Python (FastMCP client)​

Discovery, keys, and the planner​

Selected tools​

Understand tools (including Qwen3.8 and Hy4)​

Qwen-Image tools (generate / edit / VTON)​

Local dedicated VTON (Leffa + CatVTON)​

MiniMax H3 tools (video)​

NVIDIA NIM tools (understand + video)​

Google Virtual Try-On (Vertex)​

OutfitAnyone-Plus (DashScope, Beijing)​

Photoroom (Virtual Try-On / Virtual Model)​

Muse Image tools (generate / edit / VTON)​

P-Image-Ideogram (generate)​

Related​