Claude Code mcp reset-project-choices: bring back the Sume prompt

Run claude mcp reset-project-choices and Claude Code forgets the approve or reject decisions you made for servers defined in a project's .mcp.json, so the Sume entry is offered to you again. It affects project-scoped servers only; your local and user entries are untouched.

The command is listed in Claude Code's MCP docs (read 2026-10-02) as the way to reset approval choices for project-scoped servers. The Sume entry itself comes from the MCP quickstart.

Why would Claude Code not load a Sume server from .mcp.json?

Claude Code has three scopes. Local is the default and is stored in ~/.claude.json for the current project. User loads in all your projects and also lives in ~/.claude.json. Project writes to .mcp.json in the project root and is the one shared with your team through version control. Claude Code keeps approval choices for project-scoped servers, which is why a reset command exists for them.

If you rejected the Sume server once, perhaps while testing, it will not reappear on its own. The symptom is that .mcp.json clearly contains the entry, yet the server is not connected and Claude never offers it. That is the case this command is for.

How do I add Sume at project scope and reset the choice?

Add the server with the project scope flag, then reset if an earlier choice is blocking it. The URL is the production endpoint from the docs; you do not need a key for the OAuth path.

After the reset, start a session in the project and approve the server when prompted. Then authenticate with claude mcp login sume, which is the sign-in step in the Sume quickstart, and ask Claude to call tools_list to confirm.

claude mcp add --transport http --scope project sume https://mcp.sume.com/mcp
claude mcp reset-project-choices
claude mcp login sume

What does the approval mean for paid tools?

Approving a server in Claude Code is a connection decision. It is not Sume's spend gate. Sume's gates are separate: under OAuth, a session gets mcp:read by default and sees only read-only tools, and mcp:write is an explicit opt-in on the consent page. There is no mcp:paid scope, and paid calls need an idempotency_key and wallet admission, per OAuth and API keys.

That is a good reason to commit only the URL in .mcp.json and keep credentials out of it. Each teammate signs in with their own account and chooses their own Write setting. If you were considering an API key in the shared file, don't: the docs tell you to rotate keys that appear in logs or chat history, and a committed file is both.

What if it still does not show up?

Check claude mcp list for the server and its health, then open /mcp in a session. The changelog for 2.1.284 notes that /mcp no longer offers Authenticate for a server with no valid URL, so a typo in the URL looks like a missing server (Claude Code changelog, read 2026-10-02). Compare the URL character by character with https://mcp.sume.com/mcp. If the same name exists at two scopes, Claude Code does not merge them, so see two scopes with one name.

Finally, remember that the reset is a project-level convenience. It will not repair a server that fails to authenticate, and it will not change what Sume lets the session do.

Should a team commit the Sume entry at all?

Usually yes, with only the URL. A shared .mcp.json means a new teammate gets the same server name and the same endpoint, and the first session shows the approval prompt and then the sign-in. The thing to avoid is anything secret. Sume's OAuth path needs no secret in the file, and an API key belongs in an environment variable or a secret manager rather than a committed config.

Two practical habits help. First, agree on the server name; the examples in the Sume docs use sume, and prompts, permission rules and skills that mention mcp__sume__... style names only work if everyone uses the same one. Second, decide as a team whether anyone should turn Write on. Read-only is enough for browsing the catalog, listing jobs and reading assets. Generation needs mcp:write or an API key, and the spend then runs against the workspace balance. If you want a hard stop, pass max_spend_usd on paid calls and use dry_run to preview cost first, both described in MCP tools and gates. The reset command is the tool for the moment someone says they never saw the prompt; it is not a security control.