Gemini-GPT Hybrid

Multi-model orchestration subagents for Claude Code.

One model is rarely the right model for every task. Whole-codebase analysis wants a huge context window; a tight bug fix wants fast iteration. This repo packages that decision as a Claude Code subagent: it reads the scope of your request, delegates to Gemini or GPT accordingly, and routes the result back through Claude before anything reaches your codebase.

How routing works

graph TD
    A[User Request] --> B{Task Analysis}
    B -->|Large Scale| C[Gemini Engine]
    B -->|Focused Work| D[GPT Engine]
    C --> E[Claude Review]
    D --> E
    E -->|Approved| F[User Output]
    E -->|Needs Revision| G[Refinement Loop]

Signal	Routed to	Why
Whole-repo audits, architecture reviews, sweeping refactors	Gemini	1M+ token context holds the entire codebase at once
Single-file debugging, focused features, tight loops	GPT	Faster turnaround on narrow, well-scoped work
Anything that produces code	Claude review pass	Security and quality gate before output

Installation

Prerequisites

Claude Code
Gemini CLI — npm install -g @google/gemini-cli
Cursor Agent (for GPT integration)
WSL, if you're on Windows

Setup

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

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

# Windows
xcopy agents %USERPROFILE%\.claude\agents\ /E /I

Verify with ls ~/.claude/agents/ — you should see both agent files.

Usage

Mention the agent in a Claude Code conversation:

@gemini-gpt-hybrid perform a security audit on the entire codebase

Large scope — delegates to Gemini.

@gemini-gpt-hybrid debug the authentication issue in login.js

Narrow scope — delegates to GPT.

@gemini-gpt-hybrid analyze the architecture and implement improvements

Mixed scope — uses both, in sequence.

The two agents

Both live in agents/. They differ in one thing: who is allowed to write to your files.

Soft — `gemini-gpt-hybrid-soft.md`

External models analyze; Claude writes all code. Gemini and GPT never touch your files directly, so their output can't break anything on its own, and every change goes through Claude's review and your existing coding standards.

Use for production code, critical systems, and anywhere a bad write is expensive.

Hard — `gemini-gpt-hybrid-hard.md`

External models modify your codebase directly. Much faster, and correspondingly riskier — autonomous generation and sweeping refactors happen without a Claude gate in front of the write.

Use for prototypes, experiments, and throwaway code. Commit first:

git add -A && git commit -m "Backup before hard mode"
git checkout -b experimental

Hard mode can overwrite work. A clean commit beforehand is the only reliable rollback.

Choosing between them

	Soft	Hard
Writes code	Claude	External model
Review before write	Yes	No
Speed	Moderate	Fast
Rollback	Straightforward	Git only
Backup	Recommended	Required
Best for	Production	Experiments

A longer decision walkthrough lives in AGENT_SELECTION_GUIDE.md.

Review pass

In soft mode, everything an external model produces goes through Claude before it lands:

External model output
        ↓
Security audit
├─→ Injection surfaces
├─→ Authentication paths
├─→ Data exposure
└─→ Input validation
        ↓
Quality pass
├─→ Structure and cohesion
├─→ Error handling
└─→ Performance characteristics
        ↓
Approved → written to disk

This is a review layer, not a guarantee. It catches common classes of problem; it does not replace your own reading of the diff.

Contributing

Issues and pull requests are welcome — see CONTRIBUTING.md.

If you're adding an agent, follow the structure of the existing two: document the routing logic explicitly, and state which model is permitted to write.

Star history

License

MIT — see LICENSE.

Built by Seol Seunghyeon. Runs on Claude, Gemini, and GPT.