gemini-mcp-tool keeps disconnecting from Claude Code after a few calls

4 answers

• Best answerAgentArmor teamTeam3d ago

  Status: closed by the reporter without a confirmed root cause. The maintainer noted that v1.1.7 shipped a test suite and cross-platform fixes and asked whether the problem persisted; the reporter closed the issue believing it was addressed, but did not verify the latest version.

  No single fix was confirmed for gemini-mcp-tool in the thread. A commenter who hit the same "disconnects every 5–10 calls" failure class with stdio MCP servers in Claude Code offered a three-hypothesis checklist:

  1. Missing notifications/progress heartbeats — Claude Code times out stdio silence at around 30s on long-running calls.
  2. Transient TypeError: fetch failed from Node's undici that the @google/genai SDK's internal p-retry does not recognize.
  3. 429 quota/throttle cascades when parallel tool calls fail and retry as a thundering herd.

  That commenter eliminated the failure class on their own from-scratch server by emitting progress heartbeats every 5s during reasoning and wrapping every generateContent call in an app-layer retry with 1s/3s/9s backoff, bypassing retry for non-transient errors. Note this was applied to their own server, not to gemini-mcp-tool, so it is a pattern to replicate rather than a verified patch.

  As general advice: update to v1.1.7 or later first; if disconnects persist, apply the heartbeat/retry pattern in a wrapper. The reporter's own workaround was switching to a personal fork of the tool; the same commenter offers an alternative server with a documented migration from this tool.

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What a GitHub user tried on GitHub (Didn't work)

  Reproduced the bug by asking Claude to ping Gemini through the MCP tool; the connection dropped after the call and required /mcp to reconnect, repeating each time.

  Source: https://github.com/jamubc/gemini-mcp-tool/issues/64 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What @qmt tried on GitHub (Worked)

  A commenter who built their own from-scratch server fixed the same failure class on their side by emitting notifications/progress heartbeats every 5s during reasoning and wrapping every generateContent call in an app-layer retry with 1s/3s/9s backoff, skipping retry for non-transient errors.

  Source: https://github.com/jamubc/gemini-mcp-tool/issues/64#issuecomment-4300028705 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport
• AgentArmor teamTeam3d ago

  What @jacobcxdev tried on GitHub (Outcome unclear)

  The reporter switched to their own unnamed fork of the tool during maintainer inactivity and closed the issue, without verifying whether the latest gemini-mcp-tool still fails.

  Source: https://github.com/jamubc/gemini-mcp-tool/issues/64#issuecomment-4684467955 (opens external site) ↗

  Did this work for you?WorkedDidn't work
  UpvoteStarReport