JetBrains AI Assistant MCP server: add Sume with a url entry

To add Sume to JetBrains AI Assistant, open Settings | Tools | AI Assistant | Model Context Protocol (MCP), click Add, choose HTTP, and paste a JSON entry whose url is https://mcp.sume.com/mcp. Then Apply and check the Status column.

JetBrains documents this flow for Streamable HTTP servers, which is the transport Sume's hosted endpoint speaks. What the JetBrains page does not cover is how a remote server authenticates, so that part needs a test in your own IDE. This post separates what is documented from what you have to verify.

What do you paste into the New MCP Server dialog?

The JetBrains page shows a mcpServers object with one named entry and a url that points at the base HTTP endpoint. For Sume the entry is:

{
  "mcpServers": {
    "sume": {
      "url": "https://mcp.sume.com/mcp"
    }
  }
}

Where does the entry live: global or project?

The JetBrains page says you choose at the server level whether the server is available globally or only for the current project. For Sume, global is usually right because the account, balance and job tools are not tied to a repository. Pick project scope if you want a paid-capable connection to exist only in one codebase.

There is also an option to automatically enable new and changed MCP servers. Leave it on while you set up, then think about turning it off once the entry is final, because a changed entry that starts itself is the same as a changed entry that can reach a paid tool.

How do you know the connection works?

After Apply, JetBrains shows a connection state in the Status column, and clicking the status icon lists the tools the server exposes. There is also a Reconnect button for a failed start. Untick the checkbox and Apply to stop a server.

Once it is connected, ask the agent to call mcp_health, then tools_list, the discovery tools described in the Sume quickstart. The mcp_health answer includes the auth source, which tells you whether the session came in over OAuth or with a key.

What does the JetBrains page not say about authentication?

The page documents the url field only. It does not document custom headers, OAuth sign-in or minimum IDE versions for HTTP servers, so none of those are claimed here.

Sume accepts two credentials on its hosted endpoint, per the OAuth and API keys page: an OAuth access token, or an API key sent as a bearer header or x-api-key. A client that cannot send a header and cannot run an OAuth flow cannot authenticate to Sume, and nothing in the JetBrains page tells you which of those your IDE build does. Test it: if the Status column shows an error right after Apply, check the entry for typos first, then see whether your IDE offers a sign-in prompt for the server.

What can the agent do once it is connected?

Under OAuth with only mcp:read, the session sees read-only tools: jobs_list, assets_get, catalog_list and the crawl_* reads. Paid tools such as generate_image return insufficient_scope until you grant mcp:write on the consent page. With an API key the full tool set is visible, and paid calls still need an idempotency_key.

Sume does not add its own tool approval to your IDE. If you want a look before a spend, ask the agent to call a paid tool with dry_run=true first, as the tools and gates page recommends.

What does Sume not do in a JetBrains setup?

Sume's hosted MCP cannot read files from your machine, so an agent in your IDE cannot hand it a file path. For a local image or clip the flow is: create an upload URL with assets_upload_url, PUT the bytes from the client side, then call assets_complete. Each of those is a write tool, so it needs mcp:write or an API key.

Sume Image 1.0 and Video 1.0 (images_create and videos_create) are also not on hosted MCP. They stay on the REST API. The router tools generate_image and generate_video are the MCP route, and omitting payload.model sends the request to sume/auto unless you named a model family.

What should you tell the agent before it spends?

A short standing instruction helps: call tools_schema for a paid tool before using it, run dry_run=true first, pass max_spend_usd when there is a budget, and reuse the same idempotency_key when retrying a failed call. Long renders are read with jobs_wait, which answers within 55 seconds at most. If it says wait_slice_expired, the job is still running; wait again on the same id instead of submitting again.