Gemini-GPT Hybrid 🤖 Agent Open source

What it does

Gemini-GPT Hybrid is a pair of Claude Code subagents that route a request to a second model based on its scope, then bring the result back through Claude. The idea is that no single model fits every task: a whole-codebase audit wants a very large context window, while a tight bug fix wants fast, focused iteration. The subagent reads the scope of your request, delegates large-scale analysis to Gemini and narrow work to GPT (via the Cursor/Codex CLI), and routes the output back for a review pass before anything reaches your code.

What it does for your agent

• Task routing: whole-repo audits, architecture reviews and sweeping refactors go to Gemini's large context; single-file debugging and focused features go to GPT
• Two safety modes: a "soft" agent where external models only analyse and Claude writes every change after a security and quality review, and a "hard" agent where the external model writes directly for speed
• A review layer that checks external output for injection surfaces, authentication paths, data exposure and input validation before it lands on disk

Works with

Claude Code, as the host. It shells out to the Gemini CLI and to Cursor Agent / Codex for the GPT side, so you need those installed and authenticated separately. Windows users need WSL.

How to install

git clone https://github.com/NEWBIE0413/gemini-gpt-hybrid.git
cd gemini-gpt-hybrid

# macOS / Linux
cp -r agents ~/.claude/

Then mention @gemini-gpt-hybrid in a Claude Code conversation.

Maintenance and safety

MIT-licensed. It is a small two-file project rather than a large framework, and it depends on external CLIs whose behaviour and pricing it does not control. The soft mode keeps Claude as the only writer, which is the safer default; hard mode can overwrite work, and the README rightly tells you to commit first. The review pass is a filter, not a guarantee, so read the diff.

Who should use it

Claude Code users who already run Gemini and GPT CLIs and want a simple, explicit way to send big-context or fast-iteration work to whichever model suits it, while keeping a review gate in front of writes.