claude.ai MCP connectors silently stop initializing in the CLI (frozen since a version update)

Open 💬 0 comments Opened Jul 14, 2026 by robmathews

Bug: claude.ai MCP connectors silently stop initializing in the CLI after a version update

Summary

Since 2026-06-16, the Claude Code CLI no longer initializes the claude.ai-proxy
connector subsystem on startup. All six of my account's claude.ai connectors
(Gmail, github-mcp, Google Calendar, Google Drive, Linear, Slack) stopped
loading — with no connection attempt and no error logged. Manually-configured
MCP servers (HTTP/stdio in config) are unaffected and connect normally.

Net effect: connector tools such as mcp__claude_ai_Slack__slack_send_message
are unavailable in the CLI, and the connectors do not appear in /mcp, even though
they show green "Connected" on claude.ai and the feature flag is enabled.

Environment

  • Claude Code version: claude --version reports 2.1.209, but the running

process CLAUDE_CODE_EXECPATH points at 2.1.198
(/Users/rob/.local/share/claude/versions/2.1.198). Version-resolution mismatch
may be relevant.

  • OS: macOS (Darwin 25.5.0), arm64
  • Entrypoint: cli, launched top-level from a Terminal login shell
  • Auth: Claude subscription (OAuth); no ANTHROPIC_API_KEY / Bedrock / Vertex
  • Feature flag: cachedGrowthBookFeatures.tengu_claudeai_mcp_connectors = true
  • Known connectors: `claudeAiMcpEverConnected = ["claude.ai Gmail",

"claude.ai github-mcp", "claude.ai Linear", "claude.ai Slack"]`

  • Account side: all connectors show Connected at claude.ai → Customize → Connectors

Expected behavior

On startup of a top-level CLI session (subscription auth, flag enabled, connectors
authorized on the account), Claude Code brings up the claude.ai-proxy transports and
exposes their tools (e.g. mcp__claude_ai_Slack__*), and lists them in /mcp.

Actual behavior

The claude.ai-proxy subsystem is not initialized at all. No connection is attempted;
/mcp does not list the connectors; their tools are absent.

Evidence

Per-connector MCP log directories, newest log file each
(~/Library/Caches/claude-cli-nodejs/<project>/):

| Connector (log dir) | Newest log file |
|------------------------------------|------------------------|
| mcp-logs-claude-ai-Slack | 2026-06-16T12:00:13Z |
| mcp-logs-claude-ai-Gmail | 2026-06-16T12:00:13Z |
| mcp-logs-claude-ai-github-mcp | 2026-06-16T12:00:13Z |
| mcp-logs-claude-ai-Google-Calendar | 2026-06-16T12:00:13Z |
| mcp-logs-claude-ai-Google-Drive | 2026-06-16T12:00:13Z |
| mcp-logs-claude-ai-Linear | 2026-06-16T12:00:13Z |
| mcp-logs-linear (manual HTTP) | 2026-07-02T09:42:08Z |
| mcp-logs-wix-mcp-remote (manual) | 2026-07-02T09:42:08Z |

All six claude.ai connectors are frozen at the same 2026-06-16 timestamp; the
manually-configured servers logged as recently as 2026-07-02 and work today
(2026-07-14). So sessions after June 16 skip the claude.ai connectors entirely
while loading everything else.

Last-known-good connect (Slack), from mcp-logs-claude-ai-Slack/2026-06-16...jsonl:

Initializing claude.ai proxy transport for server mcpsrv_01RjgJCC9xRhSQoYQ37czqZz
Using claude.ai proxy at https://mcp-proxy.anthropic.com/v1/mcp/mcpsrv_01RjgJCC9xRhSQoYQ37czqZz
Starting connection with timeout of 30000ms
Successfully connected (transport: claudeai-proxy) in 1073ms
Connection established with capabilities: {"hasTools":true, ... "serverVersion":{"name":"Slack MCP","version":"1.0.0"}}

After 2026-06-16: no corresponding log lines exist in any session.

Steps to reproduce (as observed)

  1. Authorize claude.ai connectors (Slack, Gmail, etc.) on the account — all show Connected.
  2. Launch a top-level CLI session on the current version.
  3. Run /mcp, or attempt to use a connector tool (e.g. mcp__claude_ai_Slack__slack_send_message).
  4. Connectors are absent; MCP logs show no connection attempt to mcp-proxy.anthropic.com.

Ruled out

  • Not account/auth: OAuth subscription session, no API key; connectors green on claude.ai.
  • Not the feature flag: tengu_claudeai_mcp_connectors = true.
  • Not a config disable: no connector-gating keys in settings.json /

settings.local.json; nothing in launchd global env or shell rc files.

  • Not Slack-specific: all six connectors are affected identically.
  • Considered and discarded: CLAUDE_CODE_CHILD_SESSION=1 is present in the

session env, but it is injected into Bash-tool subprocesses (not sourced from
launchd / rc / parent shell) and was almost certainly present in the working
June-16 session too, so it does not appear causal.

Likely area

The 2026-06-16 cutoff coincides with a version change. The running-vs-resolved
version mismatch (2.1.198 executing, 2.1.209 reported) suggests the regression is in
startup version resolution and/or the claude.ai-proxy connector initialization path:
the subsystem is being skipped without emitting any log line or /mcp entry.

Requested

A silent skip is the core problem — even if there's a legitimate gating condition,
the CLI should log why the claude.ai-proxy connectors were not initialized (and ideally
surface it in /mcp) rather than omitting them with no signal.

View original on GitHub ↗