Claude-Gemini

CLI bridge for delegating heavy reasoning from Claude Code to the Gemini API. Claude is the Equipped Reasoner (tools, coordination, endurance); Gemini 3.1 Pro is the Naked Reasoner (pure deduction at ~1/7th the inference cost).

Architecture

Claude (Equipped Reasoner)          Gemini 3.1 Pro (Naked Reasoner)
+---------------------------+       +---------------------------+
| Tool orchestration        |       | Pure logical deduction    |
| File system operations    | ----> | Cross-disciplinary proofs |
| Git, bash, coordination   | <---- | First-principles analysis |
| Code implementation       |       | Structural pattern detect |
| Multi-agent management    |       | Novel problem solving     |
+---------------------------+       +---------------------------+

Commands

Command	Use Case	Default Model	Default Tier
reason	Pure reasoning dispatch (Naked Reasoner)	pro	max
analyze	Code analysis, pattern detection	auto	high
summarize	Condense large files/logs	flash	high
vision	Screenshot/image analysis	pro	high
generate	Code generation from specs	auto	high
ask	Q&A with file context	auto	high
diff	Code review of changes	pro	high
bulk	Batch process multiple files	flash	high
raw	Direct API call	flash	high
cost	Today's spending report	-	-

Deep Thinking

All commands have deep thinking enabled by default via Gemini's ThinkingConfig. Four tiers control token budgets:

Tier	Flash Budget	Pro Budget	Use Case
LOW	1,024	2,048	Simple classification, fast summarization
MEDIUM	4,096	8,192	Standard analytical tasks
HIGH	8,192	16,384	Architecture, code review (default)
MAX	16,384	32,768	Novel proofs, cryptography, combinatorial

Key Features

• Auto-routing: Automatically selects Pro for complex tasks (security, architecture, debugging) and Flash for simple ones (summarize, format, extract)
• SHA256 response caching: Identical prompts return cached results within TTL
• Daily budget enforcement: Configurable spending cap with automatic cutoff
• Retry with backoff: Handles 429/503/504 errors with exponential retry
• Stdin piping: git diff | gcop diff or cat *.py | gcop analyze --focus security
• Vision support: Multimodal image analysis (PNG, JPG, PDF)
• Structured output: All commands return JSON for easy parsing by calling agents

Setup

git clone https://github.com/herakles-dev/claude-gemini.git
cd claude-gemini
pip install -r requirements.txt

# Set your API key
export GEMINI_API_KEY="your-key"

# Or point to an .env file
export GCOP_SECRETS_PATH="~/.secrets/.env"

# Make CLI executable
chmod +x bin/gemini
ln -s $(pwd)/bin/gemini /usr/local/bin/gcop

Requires: Python 3.11+, google-genai SDK

Usage Examples

# Pure reasoning (Naked Reasoner mode)
gcop reason "Prove the logical flaw in this concurrent algorithm" --tier max
gcop reason --prompt-file problem.md --file traces.log -o solution.md

# Code analysis
gcop analyze --file main.py --focus security
cat *.log | gcop summarize --length short --tier low

# Vision
gcop vision --image screenshot.png --task ui-review

# Code review
git diff HEAD~3 | gcop diff

# Q&A with context
gcop ask --file api.py "What auth method does this use?"

# Cost tracking
gcop cost

Project Structure

bin/
  gemini            # CLI entry point (argparse, 12 subcommands)
lib/
  client.py         # Core Gemini client (generate, route_model, ThinkingConfig)
  cache.py          # SHA256 response cache with TTL and eviction
  cost.py           # Daily JSONL cost logs + budget enforcement
  prompts/
    templates.py    # Task-specific prompt templates (system + user pairs)
config/
  models.yml        # Model specs, pricing, routing rules, budget config

Tech Stack

Python, Google GenAI SDK

---

Built by D. Michael Piscitelli | herakles.dev