Making Gemini Useful With Claude

Making Gemini Useful with Claude

Use Gemini's 1M token context (free) inside Claude Desktop and Claude Code. Claude orchestrates, Gemini does the heavy lifting.

15 min setup · $0 recurring · ~200 lines of Python

Quick Start

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

Restart Claude Desktop (Cmd+Q → reopen). Test: *"Use Gemini Flash to tell me a fun fact."*
Use which python3 for the absolute path. Replace YOUR_USERNAME.

Usage

File delegation (the real power)
"Use Gemini Pro to read ~/my-project and list the 5 most complex files."

Claude sends ~50 tokens. Gemini reads the project in its 1M context (free). ~2000x token savings.

FAQ

Gemini misses context that Opus would catch?
Don't ask Gemini to judge — ask it to extract. Gemini reads 500k → returns 3k relevant code → Opus reviews those 3k. Opus on focused input beats Opus on unfocused input.

Can I use Haiku agents?
Yes. Haiku spawns → calls Gemini → gets a 200-token summary back. Cheapest Claude-side spend possible.

Slow Gemini CLI startup?
~4-8s cold start. Doesn't matter — you're using this for 15-30s tasks where 4s is noise.

Can this create images?
Not yet. Headless mode returns text only. Architecture supports it when Google adds image output.

Is this a Google ToS violation?
No. Claude never touches OAuth. The script calls Google's own CLI on your machine with your credentials. -p flag exists for programmatic use. Zero bans reported across 22k+ GitHub issues.

Why not just use shorter prompts?
Do both. Short prompts cut output ~3x. Bridge cuts input ~2000x. 650x combined.

Docs

Contributors

Thank you @Angrycomik for the Windows fix — stdin prompt passing and cwd context isolation. First contributor, first PR, shipped same day.

MIT License · Built by Ankit Dwivedi