CutPro

CutPro MCP

A Model Context Protocol (MCP) server for AI generation of every kind, AI characters and clips. It exposes the full CutPro API as tools, so an LLM can:

• Generate with any top AI model: images and videos (Veo, Kling, Seedance, Wan, Nano Banana, GPT Image), voices (ElevenLabs, Gemini), music (Suno, Eleven Music, MiniMax), sound effects, talking avatars and 3D models (Rodin, Hunyuan3D, TRELLIS), plus image and video tools (upscale, background removal, editing), from a prompt, images, a video or a sound, with each model's exact price before anything runs.
• Create AI characters: realistic people that stay consistent across photos and videos, then swap them into any video while scene, camera, motion and sound stay.
• Clip long videos and live streams: find the best moments and render the final MP4.
• Post and schedule: publish any clip, generated video or persona video to TikTok, Instagram, YouTube, Facebook, Threads, LinkedIn, Pinterest and Bluesky, now or at a set time.
• Transcribe: any audio or video into timed captions (SRT, VTT, text or JSON).

Ask in any language. Some things people do with it:

• "Make a 9:16 ad of my sneakers with Veo and schedule it on TikTok for Friday 6 pm"
• "Create an upbeat instrumental for my YouTube intro"
• "Make this photo of me say the script below in Spanish"
• "Upscale this product photo and remove the background"
• "Turn this photo of my lamp into a 3D model for my store"
• "Clip yesterday's podcast into 5 shorts with captions"
• "Faça uma narração em português deste roteiro" / "Haz un video de una cafetería al amanecer"

Key features

• Always in sync. Every v1 endpoint is a tool, generated from the live API spec: the model catalog and generation of every kind, AI characters, workspace, credits, videos and uploads, clipping, clips, channel and live monitoring, templates, renders, transcriptions, posts and connections. Chat completions and embeddings stay out: they are for code, and an AI client is already a language model.
• Runs everywhere. stdio for local clients (Claude Code, Cursor, Claude Desktop, Windsurf, VS Code, Cline, Zed) and a hosted Streamable HTTP endpoint with OAuth for ChatGPT and Claude.ai.
• Latest protocol. Speaks MCP 2026-07-28 (stateless, server/discover, cacheable lists, multi round-trip requests) and every earlier revision, on the same endpoint.
• Results you can watch. Clients that support MCP Apps show generated images, videos, voices and music, 3D files, AI characters, clips and renders as an inline player that follows a job until it is ready, with buttons to open it in CutPro, animate an image, use a voice in a video or schedule a post.
• Asks before it spends. On clients that support elicitation, every call that charges credits shows the price and waits for the user's confirmation before anything is charged.

Getting started

Requirements

• Node.js 18 or newer.
• A CutPro account (any plan) and an API key. Generate one at cut.pro/studio/me/api-keys.
• An MCP-compatible client.

Standard config

Most clients use the same JSON. Add your API key under env:

{

"mcpServers": { "cutpro": { "command": "npx", "args": ["-y", "@cutpro/mcp@latest"], "env": { "CUTPRO_API_KEY": " " } } } }

Install

After installing via a button, add your CUTPRO_API_KEY to the server's env.

Claude Code

claude mcp add cutpro --env CUTPRO_API_KEY=<your-api-key> -- npx -y @cutpro/mcp@latest

Claude Desktop

Add to claude_desktop_config.json (Settings, Developer, Edit Config):

{

"mcpServers": { "cutpro": { "command": "npx", "args": ["-y", "@cutpro/mcp@latest"], "env": { "CUTPRO_API_KEY": " " } } } }

Cursor / Windsurf / VS Code (manual)

Add the standard config above to the client's MCP settings (mcp.json / mcpServers).

Cline

Open the MCP Servers panel, choose Configure, and add the standard config above.

Gemini CLI

gemini mcp add cutpro npx -y @cutpro/mcp@latest -e CUTPRO_API_KEY=<your-api-key>

Codex

Add to ~/.codex/config.toml:

[mcp_servers.cutpro]

command = "npx" args = ["-y", "@cutpro/mcp@latest"] env = { "CUTPRO_API_KEY" = " " }

ChatGPT and Claude.ai (hosted, no install)

Use the hosted server. Add a custom connector pointing to:

https://mcp.cut.pro

You sign in to cut.pro, pick the workspace and approve the connection, so no API key and no local setup are needed. Any CutPro plan can connect; jobs use the workspace's credits. Disconnect at any time in Settings, under Connected apps.

Claude Code plugin

The plugin bundles the hosted server with a skill that teaches Claude the CutPro workflows:

claude plugin marketplace add getcutpro/mcp

claude plugin install cutpro@cutpro

Updates

Every install stays current on its own:

• Hosted (https://mcp.cut.pro, the Claude plugin and the ChatGPT app): always the latest release, nothing to do.
• Local (npx -y @cutpro/mcp@latest): on start the server checks npm and, when a newer release exists, hands the session to it. This works even for a pinned version or a stale npx cache. Set CUTPRO_MCP_AUTO_UPDATE=0 to stay on the installed version; the server then tells the model an update is available.
• Tools: new API endpoints show up as tools within ten minutes, without any release.

Every release is listed with its notes in Releases and CHANGELOG.md.

Configuration

The server is configured with environment variables.

Self-hosting the remote (Streamable HTTP + OAuth)

MCP_TRANSPORT=http PORT=8787 MCP_OAUTH=1 \

MCP_PUBLIC_URL=https://mcp.cut.pro MCP_REDIS_URL=redis://127.0.0.1:6379 \ npx -y @cutpro/mcp@latest

In OAuth mode the user signs in on cut.pro and approves the app there; cut.pro creates a key for that app and workspace, and the access token maps server side to it. Clients can register dynamically (DCR) or present a Client ID Metadata Document (CIMD). Without MCP_REDIS_URL, an in-memory store is used (single instance, state lost on restart).

Tools

The tools are built from the API's own OpenAPI spec (/api/v1/openapi.json), so they always match the API reference: one tool per endpoint, named after its operationId in snake case (analyzeVideo is analyze_video, createSubmission is create_submission), with the endpoint's own parameters, descriptions and responses. A new endpoint becomes a tool within ten minutes of reaching the API, with no new release of this package.

The usual flows:

• Anything generated: list_featured_models (or list_models for the full catalog), quote_generation for the exact price, create_generation with that price and wait, then get_generation until it is ready. A model's own fields go in options, named as in its parameters. Pass a ready image's output.image_id to animate it, or a voice's `o