agentforce314/clawcodex

What it solves

ClawCodex is a production-ready CLI agent designed for agentic coding. It solves the limitation of being locked into a single model provider by supporting a wide array of LLMs, while reducing the high token costs and noise associated with long coding sessions through advanced compression and caching techniques.

How it works

It is a Python-native rebuild of the Claude Code architecture. The agent operates in a tool-calling loop with a streaming REPL, allowing it to execute commands, fetch context, and manage session history. Key technical optimizations include:

• Token Compression (/eco mode): Uses deterministic filters to strip "ceremony" from Bash outputs (like npm or git progress) and focuses on failure summaries in test logs, reducing model-bound tokens by up to 80%.
• Prefix Caching: Maintains byte-stable request prefixes to exploit DeepSeek's prompt cache, significantly lowering costs for long sessions.
• Multi-Provider Support: Integrates 25+ providers, including OpenAI, Claude, DeepSeek, and local options like Ollama and vLLM.
• Goal-Directed Autonomy: Implements a /goal and /subgoal loop where an LLM judge verifies if the objective has been met before finishing.

Who it’s for

Developers who want a powerful, autonomous terminal-based coding agent that is flexible regarding model choice and optimized for cost and token efficiency.

Highlights

• Broad Model Compatibility: Supports OpenAI, Claude, DeepSeek, Meta, and many OpenAI-compatible backends.
• Subscription Auth: Allows using existing ChatGPT and Claude Pro/Max subscriptions instead of metered API billing.
• High Performance: Outperforms openclaude on the SWE-bench Verified leaderboard using Gemini 2.5 Pro.
• Advanced Tooling: Includes a session-scoped scheduler (/loop), MCP (Model Context Protocol) support, and a hardened permission system for tool execution.
• One-Click Install: Simple installation via a shell script for macOS, Linux, and WSL.