Claude Desktop Code tab subagents lose a "memory" MCP server

A user-configured MCP server named memory was invisible to subagents in Claude Desktop's Code tab until Claude Code 2.1.288. Sume's docs name the server sume, so a standard Sume setup never hit it, but the same release is a good reason to check that your subagents see your Sume tools at all.

What exactly was fixed?

The Claude Code changelog (read 2026-10-03) lists under 2.1.288 (October 2, 2026): "Fixed subagents in Claude Desktop's Code tab getting none of the tools of a user-configured MCP server named memory". The entry describes a name collision for one name, in one surface. It does not say other names were affected, and nothing in it concerns Sume.

What name should the Sume server use?

The MCP quickstart registers the server as sume:

• Use sume. It is the name in every Sume example, so the tool names your client shows will read like the docs.
• Some clients namespace tools by server name. Sume's MCP instructions say to use the namespaced name your client lists, so a different server name changes the tool names the agent sees.
• If a server name was already taken by another entry, pick a different name for the other one before renaming Sume.

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

How do I check that a subagent sees Sume tools?

Ask the subagent to call a read-only tool and report what it got. tools_list lists every tool visible to that session, and mcp_health shows the auth source. If the subagent says it has no Sume tools, that is a connection or name problem in the client, not a Sume permission, because a missing scope returns insufficient_scope on a tool that exists.

Does this affect the Sume Desktop connector path?

Claude Desktop can also add Sume as a custom connector by URL, which does not use a config file. The changelog entry is about the Code tab and a user-configured server, so it is a separate path. Sume does not control either client's subagent handling; it only serves the same tools to every connection.

What else should I check when a subagent has no tools?

Start from the cheapest facts. First, update Claude Code, since this release fixed a named-server case. Second, run claude mcp list in the same project and confirm sume shows as connected rather than needing sign-in. Third, check the scope the server was added under: Claude Code's MCP docs list local, project and user scopes, with local as the default, so a server added in one project does not load in another.

If your team shares a project .mcp.json, keep the Sume entry there with the URL only and let each person sign in with their own OAuth grant. Do not commit an API key. Sume's docs say to rotate a key that shows up in logs or chat history.

Subagents are a good place to use the narrower grant on purpose. A research subagent that only reads job status and catalogs can run on a read-only session, while the one that submits paid work gets Write. That split is a client decision; Sume's side is just that mcp:read sees only read-only tools.

Should I keep one Sume entry or one per project?

One user-scope entry is the least work if you always use the same workspace, since it loads in all your projects. A project-scope entry in .mcp.json makes sense when a team wants the same URL everywhere; the file holds only the URL, and each person signs in themselves. Both are described on Claude Code's MCP page, which also says local scope is the default and lives in ~/.claude.json.

Whichever you pick, keep the name sume. A rename later means every saved permission rule and instruction that mentions the old namespaced tool names has to change too, because Sume's MCP instructions tell the agent to use the namespaced name its client lists.