Sume hosted MCP is not Studio Agent: the one URL to put in your client

The only URL to give an MCP client is https://mcp.sume.com/mcp. Sume's docs are direct about the other name people mix it with: Studio Agent is the in-app Sume agent product, and it is not the public hosted MCP connector. Do not configure Studio Agent internals as a customer MCP server URL, and do not expect an agent you chat with in the Sume app to be the same thing as the tools your own client sees.

The mix-up happens because both can start a generation job and both can wait for it. They are different surfaces with different callers.

Why should it matter which one you pick? Sign-in, tools and limits differ. With hosted MCP you choose the client, the model and the prompts, and you pay through the wallet for the paid calls your agent makes. With Studio Agent, the product decides how the conversation runs.

Two surfaces, one job system

Treat the two as separate products that share the same job system underneath.

The table is short because the question is short. If the answer to "do I paste a URL into a settings screen?" is yes, you want hosted MCP, and the URL is the one at the top of this page.

Where they touch: jobs

A job belongs to its workspace and to the member whose key or Agent turn created it. That is the link between the two surfaces: a job that a Studio Agent turn started can be read from the thread it ran on, and an API key reads the jobs of its own member. The practical lesson is that you should not try to read an in-app thread's work through your MCP client unless the same member created the job. Reads for jobs of other members get 404 not_found.

The same page also says that when a turn reads a job that another member created, the callback URL and last delivery error are hidden. The details are in Jobs and results.

What to do in your client

If you want a model in Claude Code, Cursor, VS Code or Codex to use Sume tools, the work is three things: add the server URL, finish sign-in, and call tools_list once to check what the session can see. Claude Code uses claude mcp add --transport http sume https://mcp.sume.com/mcp followed by claude mcp login sume, per the quickstart. Cursor takes a url entry in its MCP config. Codex and other streamable HTTP clients take the same URL and run the client's own login.

After the connect, a tools_list call is the cheapest proof. The result depends on the grant: read-only OAuth shows read tools, a write grant or an API key shows the full hosted set, which includes generate_image, generate_video, music_create and tts_create.

claude mcp add --transport http sume https://mcp.sume.com/mcp
claude mcp login sume

Verify with one read-only call

Then ask the agent to call mcp_health. It reports the endpoint, the auth source and the safety posture, so you can see that the session is OAuth or API-key backed without guessing.

Keep the catalog in mind

Another boundary to know is the product catalog. Image 1.0 and Video 1.0 are REST-only and are not on hosted MCP; the router tools are generate_image and generate_video. The MCP overview and MCP OAuth and API keys give the full matrix of what each surface offers.

If a tool you expect is missing, catalog_list can show an HTTP capability that has no MCP tool. That is a different question from sign-in, and the answer is to use the Developer API for that capability.