Allow hiding/removing unused cloud connectors (Gmail, Google Calendar, etc.)

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/22301
2. https://github.com/anthropics/claude-code/issues/26625
3. https://github.com/anthropics/claude-code/issues/20412

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

You can disable the claude.ai cloud connectors (Gmail, Google Calendar, etc.) by setting the ENABLE_CLAUDEAI_MCP_SERVERS environment variable to false.

Add this to your ~/.claude/settings.json:

{
  "env": {
    "ENABLE_CLAUDEAI_MCP_SERVERS": "false"
  }
}

Reference: https://code.claude.com/docs/en/mcp#use-mcp-servers-from-claude-ai

Why doesn't Anthropic simply set this ENABLE_CLAUDEAI_MCP_SERVERS option to false by default? Are they stupid?

Well, their child kind of thinks so, but phrased it a bit nicer, here is full picture:

Claude Code: Disabling MCP Servers Is Not Persisted

GitHub Issues: - https://github.com/anthropics/claude-code/issues/11085 — Persistent MCP enable/disable (open since 2025-11-05, 17 comments) - https://github.com/anthropics/claude-code/issues/20412 — Cloud servers auto-injected without opt-in (open since 2026-01-23, 24 comments, 100+ upvotes) - https://github.com/anthropics/claude-code/issues/29368 — Can't hide unused cloud connectors (open since 2026-02-27) - https://github.com/anthropics/claude-code/issues/31249 — Per-server granularity for ENABLE_CLAUDEAI_MCP_SERVERS (open since 2026-03-05)

The Bug

When you disable a claude.ai-synced MCP server (Gmail, Google Calendar) via the /mcp menu in Claude Code, the disable is session-only. On the next session, the servers reappear, fail authentication, and show as failed on shutdown. There is no way to persistently disable individual cloud-synced servers.

Symptoms

Three MCP servers show as failed every time you close Claude Code
Two are claude.ai Gmail and claude.ai Google Calendar (authentication_error — no OAuth token configured)
Third is tailwindcss-docs (stderr output about callback port logged as ERROR — false positive)
Disabling via /mcp has no effect on next session
Root Cause

Claude.ai-synced MCP servers are re-fetched from the cloud on every session start. Any local disable state is overwritten. The ENABLE_CLAUDEAI_MCP_SERVERS env var only supports true/false globally — no per-server control.

Workaround Applied

Set ENABLE_CLAUDEAI_MCP_SERVERS=false in both account settings files (~/.claude-private/settings.json and ~/.claude-neo/settings.json). This disables ALL claude.ai-synced servers, not just Gmail and Calendar.

Judgment

How easy to fix: Trivial. The internal disabledMcpServers mechanism already exists for local servers. Extending it to respect cloud-synced servers is a small change — probably a single conditional check during the sync step. Adding per-server granularity to ENABLE_CLAUDEAI_MCP_SERVERS (e.g., denylist syntax) is equally straightforward.

How old: The core request for persistent MCP enable/disable is from November 2025 — over four months. The cloud-sync auto-injection complaint is from January 2026 — almost two months. Both are well-established issues with significant community engagement.

Should this be addressed: Yes. This is a basic UX expectation — disable means persist the setting. A disable button that resets every session is misleading and wastes user time. The fix is small, the demand is clear (100+ upvotes on #20412), and the workaround (nuclear kill of all cloud servers) is crude. Four months without a fix for what amounts to a one-line change is hard to justify.

Same problem on Claude.ai web, not just Claude Code on macOS.

With cloud connectors auto-loaded at the account level, baseline context on a fresh conversation starts at ~120–150k tokens before the user types anything. One tool call and I'm already brushing the 200k limit.

The per-conversation toggle (+ button → Connectors) is a manual workaround, not a fix — it doesn't persist per project. As a solo dev working across distinct project types (each needing a completely different connector set), I either:

• Disconnect at account level → painful OAuth re-dance when switching context
• Manually toggle 10+ connectors at the start of every new conversation

+1 for hiding/removing unused connectors as requested here. But the broader fix is per-project defaults — either via .mcp.json (Claude Code) or Project settings (Claude.ai), where each project declares its own connector allowlist. That would solve both this issue and the context-bloat problem in one stroke.

Thanks for the report. Same symptom as #20412. Consolidating there.

I can't believe this isn't being fixed. A couple thousand tokens getting burned for no reason. Why?