Crow is an open-source, self-hosted Model Context Protocol (MCP) platform developed by Mer@lon. It connects directly to Large Language Model (LLM) frontends and various AI clients, functioning as an agentic framework. Built with Python for its agentic components and Node.js/JavaScript for its HTTP gateway, Crow provides persistent memory, structured research tools, and encrypted peer-to-peer sharing capabilities for its users. It supports developers and professionals in building and running their own AI agents.

What does Crow connect to?

Crow functions as an MCP server platform, establishing connections with LLM frontends and various AI clients. It supports stdio transport and an HTTP gateway utilizing Streamable HTTP and SSE. Compatible AI clients include Claude Code, Claude Desktop, Cursor, Windsurf, Cline, Gemini CLI, Qwen Coder, opencode, Ollama, and any OpenAI-compatible endpoint. However, self-hosted instances do not directly support cloud web clients like claude.ai, ChatGPT, and Gemini on the web due to security considerations, with access planned for a post-v1 release.

Core features and tools

Crow exposes several callable tools for diverse functionalities. Memory API tools include list_sessions(), query_memory(query), and query_session(session_id). The built-in crow-cli agent provides tools like read, write, edit for file access, terminal for shell commands, web_search, web_fetch for web searching, capture_webcam, and read_image_file for vision input. Backend management tools are crow_register_backend, crow_list_backends, crow_remove_backend, and crow_backend_schema. The Research server offers project management with auto-APA citations and bibliography export.

Who is it for?

Crow serves a diverse audience, including developers, legal professionals, home entertainment enthusiasts, and enterprises managing regulated data. It empowers users needing persistent memory and structured research capabilities for LLM frontends, or those building AI agents capable of interacting with local systems and the web. Its encrypted sharing appeals to groups requiring secure collaboration. Skip this if direct integration with cloud web clients like claude.ai or ChatGPT is a primary requirement for a self-hosted instance.

How to get started

Access to the HTTP gateway is granted through a standard OAuth 2.1 authorization flow for remote access. For optional cloud synchronization with services like Turso, setup involves setting TURSO_DATABASE_URL and TURSO_AUTH_TOKEN environment variables.

Use cases

• Providing LLM frontends with persistent memory and structured research tools for tasks like matter organization, document research, and citation tracking.
• Enabling users to build and run their own AI agents that can read and edit code, run shell commands, and search the web, connecting natively with AI clients like Claude Code and Cursor.
• Facilitating end-to-end encrypted sharing of memories and entire project spaces between Crow users without a central server.

Crow by Mer@lon offers a self-hosted Model Context Protocol platform for AI agent development and structured research. This agentic framework supports persistent memory, tool use, and secure peer-to-peer sharing, catering to various professional and technical audiences. Its architecture allows for deep integration with AI clients, empowering users to extend AI capabilities. Try using the terminal tool to run shell commands via an AI agent.