[BUG] All claude.ai connectors stopped attaching on 2026-07-28: no claude.ai section in /mcp, no proxy transport log written since
Summary
All twelve claude.ai account connectors stopped attaching to Claude Code on 2026-07-28 and have not attached since, on any surface. They remain connected and healthy on claude.ai itself. This is not intermittent — it is a clean, total, persistent stop with a precise onset timestamp.
/mcp renders no claude.ai section at all — not an empty one, not a collapsed Show unused connectors row.
Environment
- Claude Code v2.1.223, macOS Darwin 25.5.0 (Apple Silicon)
- Auth: claude.ai subscription login (
claude_max,organizationRole: admin), no API key - Reproduces in: VS Code extension, interactive terminal CLI, and headless
claude -p
Evidence
1. Transport logs stop dead on 2026-07-28.
The last successful attach anywhere on this machine, from mcp-logs-claude-ai-Google-Drive/2026-07-27T19-59-12-714Z.jsonl:
2026-07-27T19:59:13.277Z Initializing claude.ai proxy transport for server mcpsrv_01DC4gRYi6ct…
2026-07-27T19:59:13.277Z Using claude.ai proxy at https://mcp-proxy.anthropic.com/v1/mcp/mcpsrv_…
2026-07-27T19:59:13.277Z claude.ai proxy transport created successfully
2026-07-27T19:59:13.277Z Starting connection with timeout of 30000ms
2026-07-27T19:59:13.277Z Stateless claudeai-proxy — resolving MCP initialize from cached projection
2026-07-27T19:59:13.278Z Successfully connected (transport: claudeai-proxy) in 1ms
2026-07-27T19:59:13.278Z Connection established with capabilities: {"hasTools":true,"hasPrompts":true,…}
2026-07-28T04:46:44.094Z CLAUDEAI-PROXY connection closed after 31650s (cleanly)
2026-07-28T04:46:44.094Z Cleared connection cache for reconnection
It closed cleanly and never reconnected. Since that timestamp, zero .jsonl files have been created under any ~/Library/Caches/claude-cli-nodejs/*/mcp-logs-claude-ai-*/ directory, across ~20 project directories and many dozens of sessions over 10 days. No connection attempt, no error, no timeout — nothing is logged.
2. The logging subsystem itself is fine. Local stdio servers write logs normally in the same cache tree — 15 mcp-logs-obsidian/*.jsonl files created today alone. So the missing claude.ai logs reflect a missing attach attempt, not a broken logger.
3. This install has connected successfully before. From ~/.claude.json:
claudeAiMcpEverConnected = ['claude.ai Google Calendar', 'claude.ai Gmail', 'claude.ai CMS Coverage',
'claude.ai PubMed', 'claude.ai Function Health', 'claude.ai Supabase', 'claude.ai Google Drive',
'claude.ai Uber', 'claude.ai Calendly', 'claude.ai Proofbound Auth', 'claude.ai Claude Code Remote',
'claude.ai Goodlabs']
4. Feature gates are on. tengu_claudeai_mcp_connectors: true, tengu_mcp_claudeai_eligibility_gate: true. Google Drive is not in tengu_mcp_local_oauth_blocked_hosts (that list is microsoft365/gmail/gcal only).
5. Connectors are healthy on claude.ai. Gmail, Google Calendar, Google Drive, Supabase and others all show as connected in claude.ai → Settings → Connectors.
6. All three surfaces agree.
- Interactive terminal
/mcp: sections for Local MCPs (1), User MCPs (2), Built-in MCPs (claude-in-chrome, computer-use). No claude.ai section. claude mcp list: only the 3 locally-configured stdio servers.- Headless
claude -pasked to runToolSearchwithselect:mcp__claude_ai_Google_Drive__search_files: no schema returned, so it is absent from the deferred tool registry too (not merely hidden by tool search).
Ruled out
- Auth precedence — no
ANTHROPIC_API_KEY,ANTHROPIC_AUTH_TOKEN,CLAUDE_CODE_OAUTH_TOKEN, noapiKeyHelper.oauthAccountis a valid claude.ai subscription login. - Disable lists — no
disabledMcpServers/enabledMcpServersin any user, project, or local settings file. - Stale session — reproduces in freshly launched processes.
- Stale token — ran
/login;profileFetchedAtupdated; no change in behavior. - VS Code extension — reproduces identically in a plain terminal and headless.
- Version — 2.1.223.
Expected
Account connectors attach at session start, or a visible error is surfaced when the attach step fails.
Actual
The attach step appears to be skipped entirely and silently. Nothing in any UI indicates twelve configured connectors are missing, and no diagnostic is written to disk.
Related
- #79231 — same "no log file created at all" signature, but intermittent (~3 of 20 sessions) and self-healing mid-session. This report may be the persistent form of that failure.
- #77130 —
/loginor token refresh invalidating cloud connectors in a session. - #82246 — connectors disconnecting daily in the VS Code extension.
- #39422 — closed as not planned; reporter had Gmail/Calendar working while Drive failed. Here all connectors are missing, which is a different and hopefully more tractable failure.
Happy to supply the full log tree, the mcpsrv_ ids, or run any diagnostic build.