What It Is — A Reverse-Engineered Claude Code
Formax is an open-source project that implements a Claude Code-style AI assistant for software engineering. It's explicitly inspired by Claude Code v2.0.67. The developer built it by observing network traces and behaviors, not by using upstream source code. The project is in Beta and is presented as a tool for learning, experimentation, and architecture study.
You can install it globally via npm:
npm i -g @yusifeng/formax@beta
How It Works — TUI, GUI, and a Backend Server
Formax offers multiple interfaces, mimicking Claude Code's flexibility.
1. Start in Your Project (TUI):
Navigate to your project directory and run:
cd /path/to/your/project
formax
On first run, it will prompt you for missing credentials (Anthropic or OpenAI-compatible API keys) and runtime configuration. You can also run a guided setup first with formax setup.
2. Launch the Web GUI:
After setup, run:
formax web
This starts a local web interface. A key limitation: in browser-only mode, the "Add Project" button is disabled (Desktop only). You need the Electron desktop client for full folder-picking functionality.
3. Backend for Integration:
For building your own GUI or IDE client, Formax provides a JSON-RPC backend:
formax app-server
For advanced debugging or split deployments, you can start just the WebSocket bridge:
formax serve
What's Missing — Critical Gaps vs. Official Claude Code
This is a learning tool, not a drop-in replacement. Be aware of these significant gaps:
- Incomplete Claude Code Hooks: Support for the hooks that allow Claude Code to integrate deeply with your project is still under development.
- No MCP Support: The Model Context Protocol, a core part of Claude Code's extensibility, is not supported in this version.
- Unstable Tools:
WebFetchandWebSearchhave known stability and behavior gaps. - Tool Execution Differences: The author notes that tool execution behavior is "not guaranteed to be fully identical to Claude Code."
A Unique Artifact: Built 100% with AI
The repository itself is a fascinating case study. It contains .codex/skills, docs/, and plans/ directories. The author notes these can feel noisy but are "intentionally kept as traces of AI-assisted development." If you want to understand how to build such a system with AI, opening this repo in an AI-assisted editor provides valuable context from these artifacts.
The Bottom Line: Use It to Learn, Not to Replace
⚠️ Critical Warning: The project states clearly: "Formax is experimental. Always review proposed commands and file changes before approval. You are responsible for modifications made in your environment."
At this stage, it is suited for:
- Reverse-engineering Claude Code's workflows.
- Studying the architecture of an AI coding assistant.
- Experimenting with multi-interface (TUI/GUI/Server) AI tool design.
It is not suited for:
- Stable, daily production use.
- Projects requiring reliable MCP servers or full Claude Code hook compatibility.
- Users who need the robustness and full feature set of the official Claude Code.
Provider support currently includes Anthropic and OpenAI-compatible APIs. Gemini configuration surfaces exist but runtime execution is not yet fully supported.
