MCP Tool: gemini-research — saves ~95% of research tokens

Title: gemini-research — MCP tool that saves ~95% of research tokens in Claude Code

The problem

Every time Claude Code uses WebSearch or WebFetch, it dumps 5,000 to 20,000+ tokens of raw HTML/markdown into the context window. On a research-heavy session (checking versions, looking up docs, comparing tools), you can burn through 50,000–100,000 tokens just on web searches — leaving less room for actual coding work.

The solution

I built gemini-research — an MCP server + CLI tool that delegates web research to Gemini's free API (250 requests/day with Google Search grounding). Instead of raw web pages, Claude Code gets back ~200–400 focused tokens per query.

Real-world example from today

I asked Claude Code to research GPU server rental pricing for running DeepSeek. It made 5 research calls:

| Method | Tokens consumed in context |
|--------|---------------------------|
| WebSearch/WebFetch (5 calls) | ~50,000–60,000 tokens |
| gemini-research (5 calls) | ~3,650 tokens |
| Savings | ~93% |

That's 50,000 tokens saved on a single question — tokens that stay available for actual coding.

How it works

Claude Code ──► MCP tool call ──► gemini-research ──► Gemini API (with Google Search)
                                       │
                                  check cache
                                  check rate limit
                                  build focused prompt
                                       │
Claude Code ◄── ~300 tokens ◄────── concise answer

Instead of Claude Code parsing entire web pages, Gemini does the research and returns a concise, structured summary. Results are cached locally (4h TTL) to avoid redundant API calls.

Features

• 3 MCP tools: research, check_versions, verify_github_repo
• 4 research modes: general, versions, github, deps
• Google Search grounding — real-time web data, not just LLM knowledge
• Local cache with configurable TTL
• Rate limiter tracking the 250 req/day free tier
• CLI fallback — also works via Bash: gemini-research "query" if MCP isn't loaded
• Zero cost — uses Gemini's free tier

Setup (2 minutes)

git clone https://github.com/MichaelGagnon13/gemini-research.git
cd gemini-research
pip3 install --user -e .
gemini-research --setup  # Free API key from https://aistudio.google.com/apikey

Add to your CLAUDE.md to make Claude Code prefer it:

## Research Policy
Always use gemini-research MCP tools instead of WebSearch/WebFetch.

MCP config for .claude.json:

{
  "gemini-research": {
    "type": "stdio",
    "command": "python3",
    "args": ["/path/to/gemini-research/mcp_server/server.py"]
  }
}

Why this matters for Claude Code users

• Longer sessions — context window fills up slower, fewer message compressions
• Faster responses — less data to process per research query
• Free — Gemini's free tier is generous (250 req/day)
• Drop-in — works as MCP server or CLI, no changes to workflow

GitHub: https://github.com/MichaelGagnon13/gemini-research

License: MIT — use it, fork it, improve it.

Happy to hear feedback or suggestions!