Zen4-bit/Proxima

What it solves

Proxima addresses the high cost of AI API subscriptions and the tendency for coding models to hallucinate by providing a local gateway. It allows developers to use frontier LLMs (like ChatGPT, Claude, and Gemini) inside their code editors (Cursor, VS Code, Windsurf) without needing expensive API keys, by routing queries through existing free browser sessions or local offline providers.

How it works

Proxima acts as a local development gateway running on 127.0.0.1. It offers two primary routing modes:

1. Session Routing: It emulates free accounts by routing prompts through logged-in browser sessions using sandboxed background views, capturing tokens directly from internal streams.
2. BYOK (Bring Your Own Key) Routing: It connects to official APIs (OpenAI, Anthropic, etc.) or local engines like Ollama and LM Studio, using a native keychain for secure key storage.

It integrates with editors via the Model Context Protocol (MCP) and includes a local Python agent capable of self-healing debugging, autonomous browser control via Chrome DevTools Protocol, and dynamic skill generation. To maintain context, it uses four local SQLite databases to track conversations, user preferences, experience logs (for fixing errors), and procedural skills.

Who it’s for

Developers who want to use advanced AI models in their IDE without paying for multiple premium subscriptions or high API credit costs, and those who prefer the privacy of running local models.

Highlights

• Free Browser Integration: Access frontier models via existing web sessions to bypass API costs.
• Self-Healing Agent: A Python assistant that iteratively refactors code until tests pass.
• Local Memory Engines: SQLite-backed storage for cross-session memory, factual recall, and experience learning.
• Broad Compatibility: Works with Cursor, VS Code, Windsurf, and Claude Desktop via MCP.
• Autonomous Browser Control: Ability to interact with frontends and verify UI layouts using CDP.