sundream
API & agents

Bring your own agent

Connect an MCP client to make stories, edit scenes and shots, and render media on the Sundream canvas.

Model Context Protocol (MCP) lets an external AI agent call Sundream's canvas and render tools. Changes it makes appear in the studio, and it can read work you've done in the browser.

Your agent uses its own model for writing and planning, billed by its provider. Media generated through Sundream uses your workspace's credits.

Choose a connection

Start with browser sign-in to the hosted MCP server. Connect and read your workspace for free; changes and generation require an active plan. In Settings → Connect your AI, choose your client, then approve a workspace and a daily spending limit.

ConnectionHow it works
Local bridgeRun @sundream-studio/mcp on your computer. It connects your MCP client to Sundream and can save finished media to a local folder.
Remote serverConnect a client that supports remote MCP to the endpoint in Settings → Connect your AI. Media tools return hosted URLs.

The local bridge adds syncProjectMedia when you configure a media folder. With a remote connection, you can download the returned URLs yourself.

Follow the setup guide.

Work on the canvas

The tools can make universes and stories, and edit scenes, shots, elements and boards. They can also choose models, render media, and edit the timeline of a story. Export is not in the MCP, so export a film in the studio.

Use listProjects to find the ids of universes and stories. Each tool takes the id of the place that it changes, so one connection can work on several stories.

See the tool guide for the render order, the resources and the prompt.

Set a spending policy

Each key has a render-approval mode and an optional credit cap. In confirmation mode, a render returns a quote for the agent to approve. Auto-approval starts renders directly, subject to the key's cap and workspace limits.

For several renders, an agent can open a temporary credit budget with approveSpend. See Keys and spending before starting an unattended workflow.