Vision

TheClaw is a agent runtime that inherits core principles from OpenClaw with several improvements:

1. Loose-coupled system architecture with composition of CLI commands. Which means:

• Every system capability is a CLI command.
• LLM is equiped only one bash_exec tool, with progressive discovery of system capabilities via builtin cmds CLI command.

2. Event-driven architecture with Thread (stream of events with artifacts) as first-class citizen. This will:

• Basically support agent to have persistent memory and context.
• Keep human/agent or agent/agent collaboration consistent and easily manageable.
• Improve system observability/auditability/recoverability/etc.

This Repo

TheClaw repo wraps all commands repos into one unified distributable agent platform.

NOTE: This is still in very early research and development phase, not ready yet.

Current Progress: Architecting

See TheClawArchitecture.md。

Completed:

1. pai: LLM interaction command.

• Support many providers/models, embedding, chat with basic session support and stream mode.
• It's a thin wrapper of @mariozechner/pi-ai.

2. cmds: command discover command.

• Natural language search and information for all available commands, backed by tldr-pages data and semantic search via xdb.

3. xdb: data collection command.

• Unified interface over LanceDB (vector) and SQLite (relational/FTS), with automatic embedding.
• Intent-based data collection policy for common scenarios.

4. xweb: web interaction command.

• search, fetch, explore, etc. For both humans and LLM agents.
• support brave/tavily/serper provider, fallback to simple fetch based search with Bing or Google.

5. notifier: daemon and command for task scheduling.

Planned:

1. thread: thread management command.
2. agent: agent management command.
3. xgw: gateway daemon & command.

Install

npm install -g @theclawlab/theclaw