claw-code-agent

About

The project exists for developers who want Claude Code's workflow on local models and under their own control: the entire agent runtime is standard-library Python, so the loop is auditable without node_modules, and the documented deployment is vLLM serving Qwen3-Coder with tool-call parsing configured server-side. It reimplements CLAUDE.md discovery, slash commands, session persistence, and context compaction, then extends them with MCP support, agent delegation, and cost budgets. Permission tiers default to read-only and escalate explicitly, which suits shared machines and enterprise settings. Researchers studying agent architecture and teams with data-sovereignty constraints are the primary users; the project is in alpha with active monthly feature work.