Breaking down and implementing a simplified version of how OpenClaw functions.

Manuclaw
An AI agent orchestration framework that dynamically decomposes complex user requests into executable, multi-step tool pipelines.
What it does
Manuclaw is a modular AI agent system where users interact through a terminal-based chat UI. When a user sends a request, the system uses an LLM (via OpenRouter) to automatically break it down into ordered subtasks, each mapped to a specific tool from a configurable toolset. These tools execute as a directed graph — with success/failure routing between steps — enabling autonomous multi-step workflows.
How it works

• TUI Client — A Textual-based terminal chat interface that connects to the backend over WebSockets.
• Gateway — A WebSocket server bridging the client to the AI orchestration layer.
• Subtask Generator — An LLM-powered task planner that reads a tools.json config and produces structured execution plans (tool → tool → tool) with zero manual wiring.
• Memory Module — A per-user SQLite store that logs every prompt, tool invocation, response, and status code — giving the agent persistent context across interactions.
• Tooling Config — A declarative JSON format where you define tools (input/output schemas, logic steps) and their execution flow as a directed acyclic graph.
  Example pipeline (included)

  User pastes a YouTube link → youtube_link_detection_tool extracts the video ID → youtube_transcript_fetch_tool pulls the transcript → transcript_summarizer_tool summarizes it via LLM → result returned to user.
  All of this is planned automatically by the LLM from a single user message.
  Built with
  Python · Textual · WebSockets · OpenRouter API · SQLite · httpx

https://github.com/indie-rok/manuclaw