CraftStory MCP server
Generate talking-avatar videos from Claude, Cursor, Claude Code or any other MCP client, using the CraftStory API:
- CraftStory 2.0 - a talking video of any length from one photo plus an audio clip (script + voice, your own recording, or a custom avatar). 8-15 minutes per video.
- MiniMax H3 - a clip of up to 15 s from one photo: description-driven with generated sound, audio-driven with lip-sync, or the reference model without audio (the lines written in the description are voiced; extra photos keep a second person or a product consistent). 1-3 minutes.
You need Node.js 20 or newer, a CraftStory account on a plan with API access, and an API key (app: Account -> API Access, keys look like sk-cs-...). Generations are billed in credits exactly like in the app; failed jobs are refunded.
Config
Standard MCP configuration, identical for every client (Claude Desktop, Cursor, Windsurf, VS Code, Zed...):
{
"mcpServers": {
"craftstory": {
"command": "npx",
"args": ["-y", "@craftstory/mcp"],
"env": {
"CRAFTSTORY_API_KEY": "sk-cs-..."
}
}
}
}
Install
claude.ai, Claude Desktop and mobile: hosted connector
No key and nothing to install; connectors belong to your Claude account, so one setup covers web, desktop and mobile. In claude.ai open Settings → Connectors → Add custom connector, paste
https://mcp.craftstory.com/mcp, open the advanced settings and set OAuth Client ID to
hAlwpyaDeO0CmuEKMibZxiXxtTT09NXL (public, no secret), then sign in with your CraftStory account. The
account needs a plan with API access (Producer). The hosted
server speaks Streamable HTTP with OAuth 2.1 (Auth0) and runs the same tools as the package below; local
file paths and chat attachments are not available there: pass a public URL, or let Claude call request_file_upload and drop the photo on the one-time upload page it returns.
Claude Code
claude mcp add --scope user craftstory -e CRAFTSTORY_API_KEY=sk-cs-... -- npx -y @craftstory/mcp
(--scope user makes it available in every project; drop it to install for the current project only.)
Claude Desktop
Add to claude_desktop_config.json (Settings -> Developer -> Edit Config), then restart Claude Desktop; the server shows up under the tools icon:
{
"mcpServers": {
"craftstory": {
"command": "npx",
"args": ["-y", "@craftstory/mcp"],
"env": { "CRAFTSTORY_API_KEY": "sk-cs-..." }
}
}
}
Cursor / other clients
Same shape in ~/.cursor/mcp.json (user-level, so the key never lands in a repository): command npx, args ["-y", "@craftstory/mcp"], env CRAFTSTORY_API_KEY. In a project-level .cursor/mcp.json prefer "CRAFTSTORY_API_KEY": "${env:CRAFTSTORY_API_KEY}" and keep the real key in your shell environment.
Environment variables: CRAFTSTORY_API_KEY (required), CRAFTSTORY_API_BASE (optional, default https://api.craftstory.com/api/v1; must be https unless CRAFTSTORY_ALLOW_HTTP=1).
Upload links (hosted mode and local alike): request_file_upload(kind=image|audio|video) → one-time page for the user → wait_for_upload / get_upload. Images: file_url goes into image_url, or the upload_id into create_image(image_upload_ids). Audio: upload_id goes into create_audio_clip(audio_upload_id). Audio/video: upload_id goes into reference_upload_ids of the MiniMax H3 tools. Limits: 20 MB for photos, 100 MB for audio and video.
list_models: Models, status (a paused model answers 503), limits and priceslist_voices: Library voices; include_cloned adds your cloned voiceslist_avatars: Your custom avatars, or the scenes of one avatarcreate_audio_clip: Speech from text + voice, or upload a local recordingpreview_cost: Credit estimate for a CraftStory 2.0 videocreate_craftstory2_video: Start a CraftStory 2.0 job (photo or avatar scene + audio clips)create_minimax_h3_video, create_minimax_h3_avatar_video: Start a MiniMax H3 job (basic, or reference mode with or without audio)get_job_status: Status, percentage, failure reason, refund flagget_job_result: Full record with the signed video URL (valid 7 days)wait_for_job: Bounded polling (default 45 s, max 55 s); call again while state is runningupscale_video: New job with the upscaled result (CraftStory 2.0 720p -> 1080p, H3 2x)create_image: Nano Banana image from a prompt, optionally from up to 8 of your pictures; aspect_ratio 9:16, 16:9, 4:5, 1:1, 3:4, 4:3 (13 credits, 25 with nano-banana-pro). The result's image_url goes straight into the video toolscreate_music: Music track of 5-180 s from a description, instrumental by default (1 credit per second); a declined prompt comes back with prompt_suggestioncreate_sound_effect: Sound effect or ambience of 0.5-30 s (or auto length), optional seamless loop (1 credit per started 5 s, 2 minimum)
Plus the prompt talking_video_from_photo (script + photo) that walks the model through the whole flow.
Example
Make a portrait video of the person in /Users/me/photos/portrait.jpg saying "Welcome to our spring collection", calm gestures.
The assistant will: list_voices -> create_audio_clip -> wait_for_job(audio-clip) -> preview_cost -> create_craftstory2_video (resolution 720_1280, or 768_960 for 4:5, gestures calm) -> wait_for_job(craftstory-2) a few times -> get_job_result -> the video URL.
Long jobs: wait_for_job returns within timeout_s (default 45 s, max 55 s) plus a few seconds for the final result fetch; every request inside it is capped by the time left, so it stays under the 60 s tool-call limit of most clients. A CraftStory 2.0 video needs several calls; that is by design so agent runtimes do not time out. A CraftStory 2.0 video needs several calls; that is by design so agent runtimes do not time out.
Local files vs URLs
Photos accept image_url (JPG/PNG) or image_path (absolute path or ~/...; JPG/PNG/HEIC, up to 20 MB, uploaded as multipart). Recordings and extra references are local paths too. Anything you pass as a path is uploaded to the CraftStory API, so keep your client's tool-approval prompts on (see SECURITY.md).
Hosted server (operators)
craftstory-mcp-http serves the same tools over Streamable HTTP behind OAuth. Required env: MCP_PUBLIC_URL,
AUTH0_DOMAIN; optional MCP_AUDIENCE (defaults to <MCP_PUBLIC_URL>/mcp), CRAFTSTORY_API_BASE, PORT
(8788), MCP_ALLOWED_ORIGINS, MCP_TRUST_PROXY=1 behind a load balancer. Protected-resource metadata is at
/.well-known/oauth-protected-resource/mcp; access tokens are verified against the tenant's JWKS and forwarded
to the CraftStory API, nothing is stored. Dockerfile builds the production image.
Development
npm install
npm run build
CRAFTSTORY_API_KEY=sk-cs-... npm run smoke
npm test
Full API reference: https://api.craftstory.com/api/v1/docs/public/ and the curl walkthrough at https://api.craftstory.com/api/v1/docs/samples/curl/.
Privacy Policy
This server runs on your machine and keeps no data of its own. It sends your API key and the inputs you pass to tools
(text, photos, audio, reference files) only to the CraftStory API at CRAFTSTORY_API_BASE (default
https://api.craftstory.com), where they are processed under the CraftStory privacy policy:
https://craftstory.com/privacy/. Generated videos and uploaded files are stored in your CraftStory account and can be
deleted there or via the API; nothing is retained locally by the server. No analytics or third-party services are
called by the server itself. Questions: support via https://craftstory.com/contacts/.
License
MIT