LONGTAKE
To the studio →

The studio, driven by an agent

Longtake is a cloud AI video studio built on the Seedance models — and every account can hand the controls to an AI agent. The studio ships an MCP server (Model Context Protocol, the open standard agents speak), so Claude Code, Claude Desktop, Cursor or any other MCP client can create projects, cast characters with locked voices, launch renders and read back the results. Same account, same keys, same prices — just with an agent at the desk.

Works withAny MCP client over stdio — Claude Code, Claude Desktop, Cursor
AuthA personal agent token (cf_…), issued and revoked in the studio
BillingIdentical to the UI: your provider keys, list price, no markup

What an agent can do

Twelve tools over the studio's API — enough to run a whole shoot without opening the interface.

ToolWhat it does
studio_overviewProjects, folders, scenes, characters and assets — the whole studio at a glance
create_project · create_folder · create_sceneSet up the production structure: project → episode → scene
create_characterCast a character with a portrait and a locked voice sample
generate_mocapGenerate a four-view character sheet (Seedream / Nano Banana Pro)
upload_assetAdd reference images and audio to the project library
generate_videoLaunch a take: prompt with @mentions, model, resolution, duration, variants
take_status · scene_takesFollow renders: status, video file, last frame, token spend
cancel_takeDelete a take — a still-queued one is cancelled before it bills
estimate_costPrice a generation before running it, from the exact token formula

An agent prompt then reads like a call sheet: “Create a project called Nordic Noir, cast a detective with a gravelly voice, and shoot three 5-second takes of him on a rainy pier at 480p — tell me what each cost.”

Security model

The token is issued in the studio: account menu → API keysAgent access — MCP. It signs in as you — your projects, your provider keys, your rate limits — and nothing else: it cannot issue or revoke tokens, and revoking it in the same dialog kills the agent's access without touching your password or keys. The studio stores only a hash of it.

Billing stays yours. A render launched by an agent is billed exactly like one you click yourself: the model provider charges your own API key at list price. The subscription covers agent access; there is no separate metering and no markup.

Set it up

  1. Issue a token

    In the studio: account menu → API keysAgent access — MCP → create. Copy the cf_… value — it is shown once.

  2. Get the server

    The MCP server is a single Node file with two dependencies:

    curl -O https://longtake.studio/mcp-server.cjs
    npm i @modelcontextprotocol/sdk zod
  3. Register it with your agent

    For Claude Code:

    claude mcp add longtake \
      -e STUDIO_URL=https://longtake.studio \
      -e STUDIO_TOKEN=cf_your_token \
      -- node mcp-server.cjs

    Any other MCP client takes the same command and the same two environment variables.

Questions

Does a render launched by an agent cost extra?
No. It is billed exactly like a render you click yourself: the model provider charges your own API key at its list price, and the flat subscription already covers agent access. There is no separate agent metering.
What can the token do — and not do?
It acts as your account: your projects, your provider keys, your rate limits. It cannot issue or revoke tokens, so a leaked token is contained — revoke it in the same dialog and the agent is out, with your password and keys untouched.
Which agents work with it?
Anything that speaks MCP over stdio: Claude Code, Claude Desktop, Cursor and the growing list of MCP-capable tools. The server is a plain Node file, so it runs anywhere Node 18+ runs.
Is there a plain REST API?
The MCP server is a thin client over the studio's REST API using the same Bearer token, so the API exists — but MCP is the supported surface. The raw routes aren't a documented contract yet and may change without notice.

Before the agent can shoot, the studio needs your keys: