Gemini AI Toolkit：面向终端的多模态LLM交互工具集

Gemini AI Toolkit Overview

This is a Python封装 and CLI tool for Google Gemini models, designed for terminal-preferring developers. It supports native multimodal input (text, image, video, audio, PDF) with three interaction modes: chat, text generation, and multimodal analysis. Note: The project is currently unmaintained; official alternatives like google-genai (Python SDK) and Jules (terminal AI agent) are recommended.

Key highlights:

• Terminal-native workflow to avoid web interface pain points
• Full multimodal support for diverse file types
• Flexible API parameter control and output formats

Why Build This Tool?

Developers split into two camps: web interface users (ChatGPT/Claude) and terminal-preferring engineers. Web interfaces have critical pain points:

• Rate limits: Frequent API quota triggers
• Context loss: Cross-tab conversation breaks
• Workflow disruption: Copy-paste between browser and editor

The author, a terminal-first dev, built this tool in two weeks after Google released Gemini API (Dec2023) with native multimodal capabilities. The goal: enable full-feature Gemini interaction directly in the terminal.

Three Interaction Modes

1. Chat Mode: Interactive dialogue with context maintenance (supports /clear to reset, /exit to quit).
   • CLI: python cli.py --chat
   • Python: Chat().run()
2. Text Mode: Single-shot text generation for scripting.
   • CLI: python cli.py --text --prompt "Your prompt"
   • Python: Text().run(prompt="Your prompt")
3. Multimodal Mode: Mix local files/remote URLs (supports /upload to add files).
   • CLI: python cli.py --multimodal --prompt "Task" --files file1.jpg https://url/file2.pdf
   • Python: Multimodal().run(prompt="Task", files=[...])

Supported Models & File Types

• Models: Gemini 2.0 (recommended, supports all modalities), 1.5, and 1.0 (text-only)
• File Types: Image (jpg/png etc.), Video (mp4/mov etc.), Audio (mp3/wav etc.), Documents (txt/pdf etc.)

Fine-Grained Parameter Control

Adjust generation behavior with parameters like:

• System prompt (set assistant role)
• Max tokens, temperature (randomness), top-p/top-k (sampling)
• Stop sequences, candidate count

Output Formats

• Streaming: Real-time token output (--stream)
• JSON: Structured output for downstream processing (--json)

File Handling

• Local/URL: Supports local paths and remote URLs (auto-download & cache)
• Cache: URL files stored in .gemini_ai_toolkit_cache (auto-cleaned after session)
• Google Files API: For large files (2GB max, 20GB/project storage)

Error Handling

Robust recovery for common errors:

• 429 (rate limit): Auto-retry after 15s
• Other codes (400/403/500 etc.): Clear error messages and fix suggestions

Project Status

The tool is unmaintained now. Official alternatives:

1. google-genai: Google's official Gen AI Python SDK (supported)
2. Jules: Google's terminal-first AI coding agent (jules.google.com)

Use Cases

Even unmaintained, it's valuable for:

• Terminal workflow: Fits muscle memory of terminal devs
• Script automation: Integrate into data pipelines/CI/CD
• Multimodal experiments: Test Gemini's capabilities without frontend
• Education: Example of LLM client design

Key Takeaways

Gemini AI Toolkit embodies a philosophy: minimal friction for terminal-preferring developers. It prioritizes顺手ness over full features.

While official tools have replaced it, its legacy lies in:

• Exploring terminal-native multimodal interaction
• Demonstrating user-centric tool design for niche dev groups
• Serving as a reference for future LLM client projects