🦴 Marrow

Orchestrator designed to manage a modular ecosystem of MCP (Model Context Protocol) servers.

✨ Key Features

• Logic & Output: Prevents context-flooding by giving the LLM a dedicated reasoning field and a final_answer field for user delivery.
• FastMCP: Skills run as isolated local servers using the MCP, ensuring tools can be added or removed without touching the core.
• Orchestrator: Built on LlamaIndex Workflows, ensuring every "thought" is strictly typed to a Pydantic schema (MarrowThought).
• Self-Correction: If a tool fails, Marrow reads the Python error log, understands the failure, and tries again autonomously.
• Short-Term Memory: Retains the user's original objective across complex, multi-step tool invocations.
• Stealth Scraping: Uses Playwright (Chromium) and human-like delays to bypass basic bot protection.