Origin MCP

Address

The Origin MCP address is https://mcp.origingame.ai/mcp. It is a remote MCP server (Streamable HTTP), named origin in your client. Connecting Origin MCP and publishing games are free; generating images, audio, video and 3D is charged to your Origin balance at each model's price.

Origin MCP offers platform tools only, no model or chat tool. To have your agent use Origin's models, see Connect your agent and the model API.

Install in your client

Pick your client and follow its steps to add the address and sign in. Every client takes three steps: add the address, sign in to Origin in the browser, start using it.

The full step-by-step guides with copyable commands are on the Origin MCP page. To set up several agents at once, use the installer in a terminal, or Origin Kit for a graphical app or on Windows.

Once connected, your agent shows origin and its tools, and you can ask in plain language, for example "Use Origin to generate a 16:9 pixel-art forest level background, then a main character portrait on a transparent background".

Sign-in and authorization

Browser sign-in

Origin MCP is an OAuth 2.1 protected resource, so a client that supports remote MCP sign-in needs only the address:

1. A request without sign-in gets 401, and its WWW-Authenticate header points to https://mcp.origingame.ai/.well-known/oauth-protected-resource.
2. That metadata names the authorization server https://origingame.ai/identity/v1 and the scopes create:read and create:write.
3. The client opens the browser with PKCE; you sign in to Origin and approve.
4. The client receives a short-lived access token (valid for one hour) and renews it itself.

A call that needs a scope the grant lacks answers 403 insufficient_scope; approve the larger scope in your client rather than switching credentials.

Headless use

CI, scripts and clients without OAuth send an account API token (starts with og-, created under Access) or a device credential in the Authorization header:

claude mcp add --transport http --scope user origin https://mcp.origingame.ai/mcp \
  --header "Authorization: Bearer $OG_API_KEY"

Other clients set the same header, Authorization: Bearer og-..., in their remote server configuration. Never put the token in prompts, project files, URLs or game code. The installer and Origin Kit set up a revocable device credential only for agents that cannot sign in through the browser.

Revoke access

On the Access page, agents that signed in through the browser appear under "Connected apps" with their permissions and last use; clients set up with a device sign-in appear under "Signed-in devices". Revoking one leaves your other agents and your account API token untouched.

How the tools are organized

Origin MCP lists three discovery tools first, then every operation as its own typed tool:

• docs: browse, search and read the public guides page by page.
• search: find operations by domain, name or keyword; returns the current input schema, the effect (read, write or billable) and the related guide.
• execute: compose several operations in one bounded run with origin.call(name, args). operations lists the operations this run may call; it is not proof that the user approved them.

Clients with tool search can call an operation such as generate_image directly; other clients compose operations through execute. Both paths run the same handlers, permission checks and audit. Example execute arguments:

{
  "operations": ["deploy_list"],
  "code": "async () => await origin.call('deploy_list', {})"
}

Read an operation's actual schema with search before calling it; don't guess parameter names from a guide. Irreversible operations such as deploy_delete, deploy_discard_release and game_leaderboard_reset carry destructiveHint, so clients should confirm them with you first.

Domains

The complete, current list of operations is whatever search returns. When a domain is temporarily unavailable, search names it in unavailable instead of returning empty results.

Guides

docs({ id }) reads one of these guides; docs({}) lists them all and docs({ query: "upload" }) searches by keyword.

Long jobs

For a client that declares the MCP Tasks extension (io.modelcontextprotocol/tasks), a direct call of generate_video, generate_3d, generate_tripo_3d, generate_speech, generate_motion, deploy_create or deploy_update that starts a new job answers with a task handle instead of the job's first status.

• Poll with tasks/get every 2 to 10 seconds (follow the returned pollIntervalMs); a handle is valid for 24 hours.
• When the job is done, tasks/get returns the same result as the synchronous status call, media included.
• Clients without the extension check jobs with generate_get_tasks or wait with generate_wait_tasks (up to 30 seconds per call).
• Calls inside execute and aigcc_publish are always synchronous.

Limits

execute runs in an isolated sandbox with no filesystem, network or credentials:

Don't inline build ZIPs or large images into code; use the upload flow in the local-files guide. Stopping the sandbox does not undo a write that already started: each returned call is marked started, completed, failed or unknown, so check the state before retrying.

Legacy per-domain endpoints

/generate/mcp, /deploy/mcp, /aigcc/mcp, /knowledge/mcp and /registry/mcp on mcp.origingame.ai are deprecated in favour of /mcp. Until 2026-12-31 they keep working, and every response carries Deprecation, Sunset and a Link header pointing to https://mcp.origingame.ai/mcp; after that they answer HTTP 410 (endpoint_retired). Access tokens from browser sign-in work on /mcp only.

If your agent still has origin-assets, origin-generate, origin-deploy, origin-aigcc, origin-knowledge or origin-registry, re-run the installer: it removes the old entries it wrote and keeps only origin.

Troubleshooting

When you contact support, include the operation name, the time, the task or operation ID and the HTTP status; never include tokens, cookies or signed URLs.