Claude Code ENABLE_CLAUDEAI_MCP_SERVERS=false and your Sume entry

Start Claude Code with ENABLE_CLAUDEAI_MCP_SERVERS=false claude and it will not load claude.ai connectors, so the only Sume server in the session is the one in your own MCP configuration. That removes a source of duplicate tools and a second sign-in.

The variable is listed in Claude Code's MCP docs (read 2026-10-02) as the way to disable claude.ai connectors. Sume's setup is from the MCP quickstart and OAuth and API keys.

Why would Sume appear twice?

Claude Code reads servers from your scopes and, when you are signed in to claude.ai, from connectors you added there. If you added Sume as a custom connector on claude.ai and also ran claude mcp add for the same URL, the model can see two servers that point at one endpoint. They authenticate separately, so you can end up signed in to one and not the other, and the tool list can look doubled.

This is a setup problem, not a Sume problem. Sume's hosted endpoint is one URL, https://mcp.sume.com/mcp, and each connection authenticates on its own, so two entries can carry different grants.

How do I turn the connectors off?

Set the variable for the session or put it in your shell profile. Then confirm with claude mcp list and /mcp that only the servers you configured are present.

The 2.1.282 changelog adds one more detail worth knowing: MCP servers named anthropic-skills or claude-ai list no skills or prompts, though their tools still work. Sume is not affected unless you use one of those names, so keep calling the server sume (Claude Code changelog, read 2026-10-02).

ENABLE_CLAUDEAI_MCP_SERVERS=false claude
# inside the session
# /mcp   -> expect only the servers in your own config

# or make it permanent for your shell
export ENABLE_CLAUDEAI_MCP_SERVERS=false

Which should I keep, the connector or the local entry?

Pick one per person. The connector is convenient if you use claude.ai as well as Claude Code and want one place to manage it. The local entry is better when a team shares .mcp.json, when you script Claude Code, or when you want to choose the scope a server is added at. Sume works through either, over OAuth or an API key.

For automation, the local entry is usually easier to reason about, because it does not depend on a claude.ai account. Sume's docs say API-key remote MCP remains available for automation that does not speak OAuth, and that keys give the full hosted tool set with idempotency_key and wallet admission as the gates.

What does this change for Sume's spend gates?

Nothing. Whichever route you keep, Sume's gates are the same: mcp:read by default under OAuth, mcp:write only if you opt in on the consent page, no mcp:paid scope, and idempotency_key on every write or paid call. Removing a duplicate server does not remove a risk. It removes confusion about which connection a call used.

If you ever suspect a call ran under the other connection, check jobs_list and usage_get in Sume rather than the client. Jobs and usage are tied to the account, and the job id is the evidence you want. A related post covers the same server connected twice.

If you rely on the connector in some sessions and the variable in others, write the choice down in your team's setup notes. The failure it prevents is subtle: a prompt that says to call a Sume tool works in one session and finds two candidates in another, and the model picks whichever responds first. Running one entry per session, and confirming it with /mcp at the start, removes that variable.

Finally, remember that the variable only affects what Claude Code loads. It does not sign you out of claude.ai, and it does not touch the Sume side: your Sume account, API keys and past jobs are unchanged, and you can re-enable connectors at any time by unsetting it. If a managed policy at your organization controls which MCP servers are allowed, that policy still applies on top, as covered in the post on managed MCP configuration.