Claw Code Agent

Claw Code Agent is a Python reimplementation of the Claude Code npm agent architecture, built by HarnessLab to run entirely with local open-source models via any OpenAI-compatible API server. The project is hosted on GitHub under the HarnessLab organization and is currently in alpha status, with active development tracked against the original npm source in a public parity checklist.

What It Is

Claw Code Agent is a CLI-first agentic coding tool that replicates the full Claude Code agent flow — prompt assembly, context building, slash commands, tool calling, and session persistence — entirely in Python with zero external dependencies beyond the standard library. It targets developers who want a self-hosted, locally-controlled coding agent without relying on Anthropic's hosted service. The core runtime works with vLLM, Ollama, LiteLLM Proxy, or OpenRouter as the model backend, and recommends Qwen3-Coder-30B-A3B-Instruct as the primary model.

Architecture and Runtime Surface

The project is organized around a modular Python runtime with clearly separated concerns:

• Agent loop (agent_runtime.py): Core iterative reasoning and tool-calling loop
• Tool engine (agent_tools.py): File read/write/edit, glob/grep search, shell execution, web fetch, and notebook editing
• Plugin runtime (plugin_runtime.py): Manifest-based plugins with hooks, aliases, virtual tools, and tool blocking
• MCP runtime (mcp_runtime.py): Local MCP manifest discovery and real stdio MCP transport
• LSP runtime (lsp_runtime.py): Local LSP-style code intelligence for definitions, references, hover, symbols, and diagnostics
• Session store (session_store.py): Resumable sessions with file-history journaling and snapshot IDs
• Local web GUI (src/gui): FastAPI + vanilla JS SPA launched via python -m src.gui

The permission model uses four tiers — read-only (default), write (--allow-write), shell (--allow-shell), and unsafe (--unsafe) — to keep the agent safe by default.

Deployment Model

Claw Code Agent runs entirely locally. Users clone the repository, configure environment variables pointing to a local or remote OpenAI-compatible API endpoint, and invoke the agent via python3 -m src.main. The GUI launches a local FastAPI server on port 8765. No cloud account or API key is required when using a local vLLM or Ollama backend. For cloud access, OpenRouter is supported as an optional gateway to models from OpenAI, Anthropic, Google, and others.

Supported backends:

• vLLM (recommended, with qwen3_xml tool-call parser for Qwen3-Coder)
• Ollama (OpenAI-compatible API, tool-capable models required)
• LiteLLM Proxy (gateway to many providers)
• OpenRouter (cloud API gateway, no local model server required)

Update: April 2026 Major Release

The April 2026 update is a large feature expansion. Notable additions include:

• Interactive Chat Mode (agent-chat command) with multi-turn REPL
• Local Web GUI with dark theme, sessions browser, slash command palette, tasks/plan/memory/history/worktree/skills/accounts/remote/MCP/plugins/ask/workflows/search/triggers/teams/diagnostics tabs
• Plugin Runtime with manifest-based hooks, tool aliases, virtual tools, and tool blocking
• Nested Agent Delegation with dependency-aware topological batching and agent manager lineage tracking
• LSP Runtime for local code intelligence
• Tokenizer-Aware Context with cached backends and heuristic fallback
• Prompt Budget Preflight with auto-compact/context collapse before backend failures
• Streaming output via --stream flag
• Ollama, LiteLLM Proxy, and OpenRouter backend support added alongside vLLM

The repository reports 484 stars and 207 forks as of the last recorded update (May 2026), with the last push on April 22, 2026.

Parity and Current Status

The project tracks implementation status against the original npm Claude Code source in a public PARITY_CHECKLIST.md. Areas marked in-progress include full MCP parity beyond the current stdio transport, full slash-command parity, full interactive REPL/TUI behavior, voice and VIM modes, and editor/platform integrations. The project self-describes as alpha status.