[BUG] GitHub MCP server repeatedly disconnects during web sessions, wasting tokens with no output

Resolved 💬 4 comments Opened Apr 5, 2026 by SeedClassIntelligence Closed Jul 3, 2026

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?

Description

The GitHub MCP server integrated into Claude Code web sessions disconnects and reconnects repeatedly during normal usage. This has been happening consistently for over a week.

Impact

  • Each disconnect/reconnect cycle re-injects 40+ tool definitions (~2,000+ tokens of system overhead)
  • Multiple cycles per session compound context pressure until the session times out
  • When timeout kills the response mid-generation, input tokens are billed but zero output is delivered
  • This results in direct financial loss — paying for failed completions caused by platform infrastructure, not user behavior

Reproduction

  1. Start a Claude Code web session with GitHub MCP server enabled
  2. Work normally (file edits, git operations, Python execution)
  3. Observe system-reminder messages showing MCP tools cycling between "no longer available" and "now available"
  4. After 2-3 cycles, session becomes sluggish and eventually times out

Evidence from a single session (2026-04-05)

  • 2 full MCP server disconnects observed within one conversation
  • RemoteTrigger tool dropped separately (Disconnect #1)
  • All 40+ mcp__github__* tools dropped simultaneously (Disconnect #2)
  • Git push that failed during disconnect succeeded immediately after reconnect
  • This pattern has repeated across multiple sessions daily for 1+ week

Expected Behavior

MCP server connection should be stable throughout a session. Disconnects should not inject thousands of tokens of overhead or cause session timeouts.

Request

  1. Fix the MCP server stability issue
  2. Provide a token credit for the wasted compute over the affected period — these were platform-caused failures, not user errors
  3. Allow users to opt-out of MCP servers they don't need for a given session to reduce attack surface for this bug

Environment

  • Platform: Claude Code web
  • Model: claude-opus-4-6, claude-sonnet-4-6
  • OS: Linux
  • MCP server config: Platform-injected (no user mcp.json)

What Should Happen?

I DONT KNOW. IM JUST A CUSTOMER PAYING $200 A MONTH

Error Messages/Logs

Steps to Reproduce

I DONT KNOW

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

OPUS 1M CONTEXT

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗