[BUG] Meta Ads MCP connector consistently times out (30000ms) when accessed from Claude Code IDE extension, but works fine everywhere else

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary
The claude.ai Meta Ads MCP connector (https://mcp.facebook.com/ads) fails to connect with a consistent 30-second timeout whenever it's accessed through the Claude Code IDE extension (tested in both VS Code and Google's Antigravity IDE). The exact same account/connector works fine through:

claude.ai in a browser
ChatGPT/Codex in a browser
The Claude Desktop app (as a GUI/Electron process)
Error
MCP server "claude.ai Meta Ads MCP" connection timed out after 30000ms
Shown in VS Code UI as a generic error (referred to by the user as "error 3000"), and via claude mcp get/claude mcp list in the terminal.

Environment
OS: Windows 11 Pro (10.0.26200)
Claude Code extension versions tested: 2.1.231 and 2.1.226 (both fail identically)
IDEs tested: VS Code, Google Antigravity (both fail identically)
Other MCP connectors work fine in the same environment: Canva, Google Drive, Gmail, Google Calendar (all via claude.ai), plus local stdio MCPs (analytics-mcp, google-ads-mcp)
What was ruled out (extensive testing)
Network: DNS, TCP connect, and TLS handshake to mcp.facebook.com all complete quickly and cleanly (tested via PowerShell/curl independent of the Claude Code process). A raw curl -X POST with a JSON-RPC initialize payload gets a fast 401 Unauthorized (expected without a token) — proves the server and basic HTTP/TLS path are healthy.
Config scope: Tested MCP server registered at both "local" and "user" scope — same failure both ways.
Extension version: Rolled back from 2.1.231 to 2.1.226 — same failure.
IDE: Reproduced in a second, unrelated IDE (Google Antigravity) with the Claude Code extension — rules out VS Code specifically.
Connector state: Disconnected/reconnected the Meta Ads connector via claude.ai → Settings → Connectors. This did reset state to "Needs authentication" temporarily, but after re-authenticating through the extension, it went right back to the same 30s timeout.
Cross-app conflict: Also had the same Meta Ads MCP connected via Codex/ChatGPT — removed that connection (and removed the app from Facebook Business Integrations) to rule out a shared-session/token conflict. No change — still timed out.
Firewall: Windows Firewall outbound rule created explicitly for the extension's claude.exe binary. No change.
Debug flags: --debug doesn't appear to instrument the mcp get/mcp list subcommands — no extra logs obtained.
Root cause confirmed via debug logs
Ran /debug inside Claude Code and inspected the resulting logs in ~/.claude/debug/*.txt. Each claude.ai-managed connector goes through a dedicated proxy transport at https://mcp-proxy.anthropic.com/v1/mcp/<mcpsrv_ID>, where <mcpsrv_ID> is a fixed internal ID per connector. For this account, the Meta Ads connector's ID is:

mcpsrv_01XjpwMigrPks2YzhKUzWzSV
The log shows the connection to this specific proxy endpoint is opened and then simply receives zero bytes back for exactly 30001ms — no success, no HTTP error, nothing — until the client gives up:

Connection timeout triggered after 30001ms (limit: 30000ms)
In the same log/session, other connectors going through the same mcp-proxy.anthropic.com host (different mcpsrv_ IDs) — Canva, Google Drive — connect successfully in 27–1663ms. So the proxy mechanism itself works; it's this one mcpsrv_ ID that never responds.

This rules out the client environment entirely. The client never even reaches mcp.facebook.com — it's stuck waiting on Anthropic's own mcp-proxy.anthropic.com for this specific connector ID. All earlier local troubleshooting (DNS/TCP/TLS to Facebook, firewall rules, extension version, IDE choice, reconnecting the connector, removing the Codex/ChatGPT integration) is consistent with this: none of it could have mattered, because the hang happens one hop earlier, server-side, before Anthropic's proxy ever talks to Meta.

Why this looks isolated to the Meta Ads connector specifically
In the same claude mcp list output, in the same session, all other claude.ai-proxied connectors (Canva, Google Drive, Gmail, Google Calendar) connect successfully (✔ Connected) — only Meta Ads times out.

Note on Claude Desktop app working
At one point, testing from the Claude Desktop app (GUI/Electron process) against this same Meta Ads connector succeeded. Given the confirmed root cause is server-side (specific to this mcpsrv_ ID on mcp-proxy.anthropic.com), that success was most likely a transient backend state, not a reproducible client-side difference — please don't read too much into the Desktop-vs-CLI distinction; it may not be meaningful now that the real cause is known.

Request
This looks like a backend bug in mcp-proxy.anthropic.com specific to connector ID mcpsrv_01XjpwMigrPks2YzhKUzWzSV (the Meta Ads connector on this account) — some stuck/misrouted state that never gets a response from whatever it proxies to. Would appreciate someone investigating that connector ID server-side; this isn't something fixable from the client/OS side.

What Should Happen?

connect status sucess

Error Messages/Logs

MCP server "claude.ai Meta Ads MCP" connection timed out after 30000ms

Steps to Reproduce

Steps to Reproduce
In claude.ai → Settings → Connectors, connect the "Meta Ads MCP" connector (mcp.facebook.com/ads) and complete the Meta/Facebook OAuth login successfully.
Open the Claude Code extension inside an IDE (reproduced in both VS Code and Google's Antigravity), in a project that has this connector available.
Run claude mcp list (or claude mcp get "claude.ai Meta Ads MCP") from a terminal, or simply try to use a Meta Ads MCP tool from within an agent session.
Observe: the connector consistently fails after exactly ~30 seconds with MCP server "claude.ai Meta Ads MCP" connection timed out after 30000ms. Other claude.ai connectors (Canva, Google Drive, Gmail, Google Calendar) succeed normally in the same mcp list run.
For confirmation of where it hangs: run /debug inside a Claude Code session and inspect ~/.claude/debug/*.txt — the log shows the request to https://mcp-proxy.anthropic.com/v1/mcp/mcpsrv_01XjpwMigrPks2YzhKUzWzSV (this connector's fixed proxy ID) receives zero response for 30001ms before the client aborts.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.229

Claude Code Version

2.1.229

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_