Three Popular Open‑Source AI Coding Projects: CodeWhale Agent, Free Claude Code Proxy, and Gemini‑Web2API

First: CodeWhale — 36K+ Stars Terminal AI Coding Assistant

CodeWhale is a terminal programming proxy written in Rust that connects to DeepSeek V4 and MiMo models and wraps them with a permission system called “Constitution”. It offers three modes: Plan (read‑only), Agent (file operations require approval), and YOLO (full autonomy).

Core Features

Constitution system : nine‑level authority hierarchy hard‑coded in the binary; higher‑priority user commands override outdated project configs, and tool output overrides model guesses. It leverages DeepSeek V4 prefix caching, keeping cold‑start cost around 1%.

Three‑mode control : Plan / Agent / YOLO with clear permission granularity; each file operation can be whitelisted or blacklisted.

Automatic snapshots : creates a git snapshot after each operation round, enabling one‑click rollback.

Parallel agents : can run up to 20 sub‑agents concurrently; after code edits it automatically runs language‑specific LSP diagnostics (rust‑analyzer, pyright, typescript‑language‑server, etc.).

Installation & Usage

Four installation methods are supported:

# npm one‑click install
npm install -g codewhale
# Homebrew
brew tap Hmbown/deepseek-tui && brew install deepseek-tui
# Docker
docker pull hmbown/codewhale
# Cargo (requires Rust 1.88+)
cargo install codewhale

Notes

The Constitution is essentially a very long prompt; whether the model strictly follows the nine rules varies by user. Rust compilation can be slow, but npm installation is unaffected.

Project link: https://github.com/Hmbown/CodeWhale<br/>Stars: 36K+ | Language: Rust | License: see repository

Second: free‑claude‑code — 31K+ Stars, Zero‑Cost Claude Code

Claude Code is Anthropic’s official terminal coding tool, which normally requires a paid subscription. free‑claude‑code provides a simple proxy that routes Claude Code API calls through alternative model back‑ends, eliminating the subscription cost.

Supported Models

Supports 17 back‑ends, including NVIDIA NIM, OpenRouter, Google AI Studio (Gemini), DeepSeek, Mistral, Groq, LM Studio, Ollama, etc.

Three model tiers can be routed separately: Opus (expensive), Sonnet (cost‑effective), and Haiku (free). For example, use Gemini 2.5 Pro for Opus‑level tasks, DeepSeek V3 for Sonnet, and a free model for Haiku.

Usage

curl -fsSL "https://github.com/Alishahryar1/free-claude-code/blob/main/scripts/install.sh?raw=1" | sh

After installation, start Claude Code with the command fcc-claude. The web management UI is available at http://127.0.0.1:8082/admin, where API keys and model routing are configured.

Extended Features

Provides Discord and Telegram bot wrappers for remote task submission, and integrates Whisper and NVIDIA NIM for voice‑to‑text, enabling voice input of programming requirements.

Notes

Each model requires its own API key. Gemini’s free quota has a daily limit; NVIDIA NIM imposes rate limits. Claude Code’s deep reasoning and parallel calls behave inconsistently across models, so switching models may change the experience.

Project link: https://github.com/Alishahryar1/free-claude-code<br/>Stars: 31K+ | Language: Python | License: MIT

Third: gemini‑web2api — Turn Gemini Web into a Free API

Google Gemini’s web UI is free but lacks a public API. gemini‑web2api acts as a local proxy that masquerades Gemini Web’s endpoints as OpenAI‑compatible, allowing any tool that supports the OpenAI API (e.g., Cherry Studio, ChatBox, OpenAI Python SDK, Codex CLI) to use Gemini as a backend without cost.

Feature Description

Supported models include Gemini 3.5 Flash, Flash Thinking (supports ultra‑long output), Pro, Auto, Lite, etc. Adding @think=N after the model name adjusts thinking depth (0 = deepest, 4 = shallowest).

Three interaction types are available: chat (conversational), thinking (deep reasoning), and auto (automatic mode switching).

Usage Steps

The implementation is pure Python with no third‑party dependencies: python gemini_web2api.py By default it listens on http://localhost:8081/v1. Point any OpenAI client’s base URL to this address to use Gemini. Anonymous mode works out‑of‑the‑box; using Pro models requires providing a cookie.

Notes

This project is essentially reverse‑engineered web‑interface code; Gemini’s web endpoints may change at any time, and cookies can expire. Free quota is subject to official rate limits, making it unsuitable for high‑traffic scenarios. Use only for learning and research.

Project link: https://github.com/Sophomoresty/gemini-web2api<br/>Stars: ~400 | Language: Python | License: MIT