Bug: MCP Tool Response Exceeds Token Limit (45k tokens) Even for Small Prompts

[tjazevedo]

Bug Report: MCP Tool Response Limit - Model-Specific Issue

Problem Description

The MCP tool gemini-cli is returning a token limit error with specific models, even for very small prompts:

Error: MCP tool "ask-gemini" response (45735 tokens) exceeds maximum allowed tokens (25000)

Issue Details

• Configured limit: 25,000 tokens
• Actual response: 45,735 tokens (always the same number)
• Issue: Occurs regardless of prompt size
• Behavior: Even 10-word prompts generate 45k+ token responses

Model-Specific Analysis

After thorough testing, this is a model-specific bug:

Model	Status	Behavior
gemini-2.5-pro (default)	❌ BROKEN	Always returns 45,735 tokens
gemini-2.5-flash	✅ WORKING	Normal response sizes
gemini-2.0-flash-thinking	❌ 404 ERROR	Model not found

Reproduction Steps

1. Use gemini-cli MCP tool with default model (gemini-2.5-pro):

   /gemini-cli:analyze "What is 2+2?"
   

   Result: 45,735 token error

2. Use gemini-cli MCP tool with flash model:

   /gemini-cli:analyze -m gemini-2.5-flash "What is 2+2?"
   

   Result: Works perfectly (normal response size)

Environment

• Tool: gemini-cli MCP tool v1.1.1
• Context: Claude Code interface
• Node.js: v20.19.3
• Google Gemini CLI: v0.1.10
• Configuration: Properly configured with claude_desktop_config.json

Installation Verification

All installation requirements are met:

• ✅ Node.js ≥ v16.0.0 (have v20.19.3)
• ✅ Google Gemini CLI installed and configured
• ✅ MCP server configured correctly via NPX method
• ✅ claude_desktop_config.json properly set up

Expected Behavior

• Response should respect the 25,000 token limit
• All models should work consistently
• Large responses should be truncated or paginated

Workaround

Use gemini-2.5-flash model explicitly:

/gemini-cli:analyze -m gemini-2.5-flash "your prompt here"

Root Cause Analysis

This appears to be a model-specific bug in gemini-2.5-pro:

• The model returns a hardcoded response size of 45,735 tokens
• This happens regardless of the actual prompt content
• The issue is in the Gemini API response for the Pro model, not the MCP tool itself

Suggested Solutions

1. Fix gemini-2.5-pro model to return appropriate response sizes
2. Implement model-specific token limits in the MCP tool
3. Add automatic fallback to gemini-2.5-flash when Pro model fails
4. Update documentation to recommend using Flash model for analysis tasks
5. Add model validation to prevent using broken models

Additional Context

• This is not an installation issue - all components are properly configured
• The MCP tool works perfectly with compatible models
• The bug is isolated to the gemini-2.5-pro model specifically
• Logs show MCP server initializes correctly

Related

• Consider making gemini-2.5-flash the default model
• Add model health checks before processing requests
• Update troubleshooting documentation with model-specific issues

Update Status

Installation verified as correct - Issue is confirmed as model-specific bug in gemini-2.5-pro.

[jamubc]

does using "--debug" flag with claude code: claude --debug give any clues?

[akasimo]

I am experiencing the same last few days, geminis answers are too long for claude. And for me it doesnt make sense to dumb down gemini, would rather use pro instead of flash in most situations

[jamubc]

Caching should have resolved this problem, although I need to test further.