Per-project disable for claude.ai-connected MCP integrations

Claude.ai-connected MCP integrations (Atlassian, Figma, Gmail, Google Calendar, Notion, Sentry, Drive, Excalidraw, Microsoft Learn) are activated globally as soon as they're connected at the account level. There is no per-project, per-workspace, or per-session way to disable a specific integration without disconnecting it from the account entirely.

Why this matters

When many integrations are connected, the resulting tool count pushes other MCP servers (in my case cli-bridge, which exposes a domain-specific Go indexer) into the deferred tool bucket. Deferred tools require an extra ToolSearch round-trip before use, which agents systematically skip in favor of trained reflexes (Read, Bash, Grep). This neutralizes a lot of the value of building specialized MCP tooling.

In a controlled eval (one MCP server, no claude.ai integrations), my custom tools were picked up at high rates. In daily use (10 integrations connected), they almost never are.

Requested behavior

Either:

1. Extend disabledMcpjsonServers (or add a sibling setting) to accept claude.ai integration IDs as well as local MCP names, scoped at user, project, and .mcp.json levels. Or
2. Add an "active per project" toggle in the Claude.ai Connected Apps settings, so an integration can stay OAuth-connected but only surface in selected workspaces. Or
3. At minimum: opt-in defaults — newly connected claude.ai integrations should not auto-activate in every Claude Code workspace; require explicit per-project enable.

Option 1 is the most useful for power users. Option 3 is the smallest behavioral change with the largest UX win.

Workaround in the meantime

Disconnect the integrations at claude.ai Settings → Connected apps. Reconnect when needed. This is annoying because OAuth round-trips are visible UX; the local schema-suppress is invisible.