VideoGen

@videogen/mcp

A Model Context Protocol (MCP) server that exposes the full VideoGen API to any MCP client (Cursor, Claude Desktop, Windsurf, etc.). Your agent can generate videos from scripts, produce images / voiceovers / music / avatars, upload files, export and remix projects, and manage runs — all authenticated with your own API key.

It ships in two transports. Tool surface is the same except for ChatGPT Apps commerce policy (see below):

• Local (stdio) — the client launches @videogen/mcp as a subprocess and reads the key from VIDEOGEN_API_KEY.
• Remote (Streamable HTTP) — a hosted, multi-tenant HTTP server. Clients connect over the network and authenticate per request with Authorization: Bearer <api-key>; no key lives on the server.

Host surfaces (remote): /mcp (and stdio) is the STANDARD surface: when the user is out of credits or needs a plan feature, tools and guidance walk them into get_app_deep_link (OPEN_UPGRADE / OPEN_PURCHASE_CREDITS / OPEN_ENABLE_TOP_UPS). /mcp/chatgpt is the ChatGPT Apps surface: those commerce deep links are omitted, billing errors are rewritten to “manage your VideoGen account,” and copy never says purchase / buy / upgrade / top-ups (OpenAI Plugins digital-goods policy). See .cursor/rules/chatgpt-mcp-no-commerce.mdc and mcp/src/hostSurface.ts.

This is distinct from the hosted documentation MCP at https://docs.videogen.io/_mcp/server, which only lets clients read the API docs. This server actually executes the API.

How it works

• Wraps the official @videogen/sdk and calls the live API through the SDK's authenticated passthrough.
• The local transport reads your key from VIDEOGEN_API_KEY; the remote transport reads it from each request's Authorization: Bearer header and builds a fresh, per-request server bound to that team (stateless — no session state is shared between requests).
• The key never leaves your machine except in requests to the VideoGen API (local), or is forwarded only to the VideoGen API for the duration of the request and never persisted (remote).
• Long-running operations (workflows, media tools, project exports) use composite tools. The hosted HTTP transport returns the run/execution id immediately; use the corresponding get_* tool to continue polling. The local stdio transport waits for completion by default.

Configuration

Get an API key from app.videogen.io/api. Both transports expose the same tools; the remote server is recommended.

Remote (Streamable HTTP) — recommended

Nothing to install or update. Point any MCP client that supports the Streamable HTTP transport at the hosted endpoint and send your API key as a bearer token:

{  "mcpServers": {    "videogen": {      "url": "https://mcp.videogen.io/mcp",      "headers": {        "Authorization": "Bearer sk_videogen_live_..."      }    }  }}

The remote server:

• Accepts MCP JSON-RPC messages via POST /mcp (stateless — a fresh server per request).
• Reads the API key or OAuth access token from the Authorization: Bearer header. The standard /mcp endpoint requires credentials for every request and returns 401 with a WWW-Authenticate challenge to start OAuth. The ChatGPT-specific /mcp/chatgpt endpoint serves tool discovery unauthenticated and returns its OAuth challenge on a protected tool result because ChatGPT does not start OAuth from the standard transport challenge. The credential is forwarded only to the VideoGen API and never stored.
• Exposes GET /health for load-balancer / Cloud Run startup probes.
• Handles CORS preflight (OPTIONS) so browser-based clients can connect.
• Supports three upload paths. For small assets (images, logos, short audio), upload_file takes base64-encoded contents inline (fileData). For large files, create_file_upload returns { fileId, uploadUrl }; the client PUTs the raw bytes to that short-lived pre-signed URL (no Authorization header) and then calls get_file with { fileId, wait: true } to wait for processing. In ChatGPT (an MCP Apps host), open_uploader renders an in-chat upload widget so the user can pick a file directly: the widget itself calls create_file_upload, PUTs the bytes client-side, and reports back only the resulting vg_file_... id, so the pre-signed URL is never surfaced to the model. Either way you get a vg_file_... id to pass to other tools. The local (stdio) server instead uploads by local filePath. (The server never fetches a caller-supplied URL, so there is no SSRF surface.)

Local (stdio) — Cursor / Claude Desktop

Runs as a subprocess launched by your client with npx. Reads the API key from the VIDEOGEN_API_KEY environment variable, which never leaves your machine except in requests to the VideoGen API:

{  "mcpServers": {    "videogen": {      "command": "npx",      "args": ["-y", "@videogen/mcp"],      "env": {        "VIDEOGEN_API_KEY": "sk_videogen_live_..."      }    }  }}

Environment variables

Tools

Workflows (end-to-end video)

script_to_video, voiceover_to_video, slideshow_to_video, storyboard_to_video, list_workflow_runs, get_workflow_run, cancel_workflow_run

Media tools

generate_image, generate_video_clip, text_to_speech, generate_sound_effect, generate_music, generate_avatar, vectorize_image, remove_image_background, remove_video_background, upscale_image, upscale_video, image_3d_effect, list_tool_executions, get_tool_execution, cancel_tool_execution

Projects

list_projects, get_project, export_project, get_project_export, remix_project, list_project_remix_actions

Files

upload_file, create_file_upload, get_file, list_files, open_uploader

open_uploader is a ChatGPT App widget (remote server only): it renders an in-chat file picker (a React component served as an MCP UI resource) so a ChatGPT user can attach a file without pasting a link. It is a no-op on clients that don't render MCP Apps UI — those use upload_file / create_file_upload instead.

Entities

list_entities, create_entity, get_entity, update_entity, archive_entity, add_entity_reference, remove_entity_reference

Create ACTOR / PRODUCT / VISUAL_STYLE entities, attach uploaded image references, then pass vg_enti_... ids into workflows and generate_avatar (actorEntityId, storyboard entity attachments, etc.).

Resources & account

list_tts_voices, list_languages, get_me, get_app_deep_link

Guidance (docs for agents)

Operational manuals exposed as MCP resources and mirror tools (no API credential required):

Call the matching get_*_guidance tool before non-trivial setup, polling, workflow/remix/export, or tools-vs-workflows decisions. Many hosts never auto-attach resources; the tools are the reliable path. Content is grounded in the public docs at docs.videogen.io but written for MCP tool usage (including hosted wait caps).

Creative MCP tools expose user intent rather than the lower-level developer API request shape. Use style for a full, strict visual-style paragraph (medium, texture, palette, then a simple composition lock) and aspectRatio ({ width, height } units, e.g. { width: 16, height: 9 }) for output dimensions. Omitted workflow styles use the app Realistic look (Photorealistic photograph, natural lighting). Do not pass a short label such as watercolor. Image models pack the frame with text, charts, and diagrams unless the style keeps the picture simple. remix_project accepts curated edits: CAPTIONS, TRANSITIONS, CONVERT_IMAGES_TO_VIDEOS, and ZOOM. CONVERT_IMAGES_TO_VIDEOS generates AI video clips from stills (expensive). ZOOM is cheap Ken Burns camera motion. watermarkMode and endScreenMode are not exposed; MCP always sends AUTO (Free-plan output includes the VideoGen watermark; VideoGen Pro removes it).

Workflow, media-generation, and export tools start the operation and poll internally. On the hosted (Streamable HTTP) server, the wait window is capped under Cloudflare's proxy timeout, so long generations return a still-running snapshot instead of a 524. Use get_tool_execution, get_workflow_run, or get_project_export to continue polling.

For generate_avatar, provide audioFileId and an ACTOR entity via actorEntityId. You may set avatarQuality to LOW, STANDARD, HIGH, or MAX. script_to_video, slideshow_to_video, and CHANGE_NARRATOR accept the same optional actor fields.

See the full tool reference for every tool's parameters and its REST-endpoint mapping.

Development

pnpm build           # bundle to dist/ (dist/index.js, dist/http.js, and the widget)pnpm typecheck       # type-checkpnpm lint:fix        # lint
# Smoke test the local (stdio) server:VIDEOGEN_API_KEY=sk_videogen_live_... node dist/index.js
# Smoke test the remote (HTTP) server:PORT=8080 node dist/http.js#   Health:  curl http://localhost:8080/health#   MCP:     curl -X POST http://localhost:8080/mcp \#              -H 'Authorization: Bearer sk_videogen_live_...' \#              -H 'Content-Type: application/json' \#              -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

VIDEOGEN_BASE_URL overrides the upstream API for both transports (e.g. point at a local API during development).

Deployment (remote server)

The remote server deploys to Cloud Run through the standard monorepo CI/CD, alongside api / backend / external / frontend:

• Image: ci-cd/docker/mcp/Dockerfile — a minimal multi-stage Node build (no ffmpeg / gcloud / VPC tooling; the server only makes outbound HTTPS calls to its per-environment VideoGen developer API — dev.api.videogen.io / prerelease.api.videogen.io / api.videogen.io, all of which exist).
• Service registry: mcp is registered in ci-cd/utils.ts (CLOUD_RUN_SERVICES).
• Build config: ci-cd/config.build.ts (CLOUD_RUN_SERVICE_TO_BUILD_CONFIG_MAP.mcp).
• Deploy config: ci-cd/config.deploy.ts (mcp) — 1 CPU / 2 GiB, concurrency 80, 60m request timeout (tool calls long-poll workflows/exports), allowUnauthenticated: true (auth is per-request via the caller's API key, not GCP IAM).
• Generated pipelines: ci-cd/cloudbuild/mcp.build.yaml and the mcp entries in the GitHub build/deploy workflow matrices are produced by the CI/CD generators — regenerate them (do not hand-edit) after changing the build/deploy config.

Manual GCP steps (one-time)

CI/CD builds and deploys the Cloud Run service, but a couple of steps must be done by hand in GCP the first time:

1. Custom domain / DNS. The service is reachable at its generated *.run.app URL immediately. To serve it at mcp.videogen.io, create a Cloud Run domain mapping (or add it behind the existing load balancer) and add the corresponding DNS record. Update the url in the client config above once the domain resolves.
2. Verify the first deploy. After the first successful deploy, confirm GET https://<service-url>/health returns 200 and that a POST /mcp with a valid bearer token lists tools.

ChatGPT connector (OAuth) checklist

After deploying the remote MCP server for an environment (e.g. DEV → https://dev.mcp.videogen.io/mcp):

1. Confirm discovery + auth modes:
   • GET https://<mcp-host>/.well-known/oauth-protected-resource/mcp returns 200 with resource ending in /mcp and authorization_servers equal to the MCP origin (pathless — Cursor workaround).
   • GET https://<mcp-host>/.well-known/oauth-protected-resource/mcp/chatgpt returns 200 with resource ending in /mcp/chatgpt and authorization_servers equal to the Supabase Auth issuer (…/auth/v1).
   • GET https://<mcp-host>/.well-known/oauth-authorization-server returns 200 with issuer equal to the MCP origin and authorize/token/register endpoints on Supabase Auth.
   • npx -y @modelcontextprotocol/[email protected] --cli https://<mcp-host>/mcp/chatgpt --transport http --method tools/list lists ~35+ tools (including open_uploader as noauth and API tools as oauth2).
   • Every anonymous POST against /mcp, including initialize and tools/list, returns HTTP 401 + WWW-Authenticate so Cursor / Claude start OAuth immediately.
   • Anonymous tools/call get_me against /mcp/chatgpt returns HTTP 200 with a tool-result _meta["mcp/www_authenticate"] challenge (ChatGPT).
2. In the ChatGPT app (e.g. VideoGen (DEV)), set the MCP URL to https://<mcp-host>/mcp/chatgpt (not /mcp). Copy the exact OAuth callback URL (https://chatgpt.com/connector/oauth/{callback_id}) into that environment's Supabase Auth OAuth client redirect allowlist.
3. Prefer DCR in the ChatGPT connector builder (Supabase advertises registration_endpoint; it does not advertise CIMD / client_id_metadata_document_supported). After ChatGPT DCR's a client for our app, add that exact client_id to TRUSTED_OAUTH_CLIENT_IDS in core/src/logic/oauth/trustedOAuthClientIds.ts so the consent screen treats it as verified.
4. ChatGPT Settings → the app → Refresh → Scan Tools → complete the OAuth prompt when shown.
5. Start a new conversation, select the app from the tools menu, and call a read tool (e.g. get_me). Do not reuse an old chat after metadata changes.

STANDARD_TESTS runs anonymous ChatGPT discovery against /mcp/chatgpt automatically via mcp-http (pnpm --filter @videogen/api mcp:test -- --transport http --env <env> --server-mode remote), which invokes the Inspector CLI before the authenticated full smoke and asserts open_uploader advertises noauth and API tools advertise oauth2. That scheme check fails against a host that has not yet been redeployed with this metadata.