ChatCrystal

What ZengLiangYi/ChatCrystal MCP server does

ZengLiangYi/ChatCrystal MCP server organizes coding conversations from supported local tools into a personal knowledge base. Import sources include Claude Code, Cursor, Codex CLI, Trae, and GitHub Copilot. Imported conversations can be distilled into notes containing titles, summaries, conclusions, code snippets, and tags.

The resulting knowledge base supports semantic search, tag-based browsing, and a tag knowledge graph. Search uses embeddings and can expand results through relationships between related knowledge points. Markdown export is also available for taking notes outside the application.

The MCP interface lets agents recall stored experience and write reusable experience back to the knowledge base. The project also provides a desktop application, a CLI, and a web interface, so MCP is one access method within a broader local workflow.

How it works

ChatCrystal scans local tool data directories during import, rather than requiring conversations to be manually copied into the application. Summarization requires an LLM provider, while semantic search requires a separate embedding provider. These providers are configured independently; a language model is not treated as an embedding model.

The application runs locally by default and stores its working data locally. Its CLI can start the service, import conversations, search notes, list or inspect notes, run batch summarization, view configuration, and start the MCP server. The MCP entry point runs over standard input and output with crystal mcp.

ZengLiangYi/ChatCrystal MCP server can also be deployed with Docker for access from multiple devices. In that mode, the service uses a persistent chatcrystal-data volume and can be protected with a shared API token. Provider URLs can be configured through the web UI or Docker-related settings.

Setup and configuration

The project requires Node.js 24 or newer. The npm package can start the MCP server directly:

npx -y chatcrystal mcp

For the broader CLI and web workflow, install the package globally, start the service, and open the local web interface:

npm install -g chatcrystal
crystal serve -d
crystal import

After installation, configure both an LLM provider and an embedding provider. The README identifies OrcaRouter as a supported OpenAI-compatible option for LLM generation, and also describes using Ollama or external APIs. The desktop application provides provider settings before importing data.

Limitations and notes

The repository lists Windows as its platform and requires Node.js 24 or newer. Semantic search is not available without an embedding provider, and note summarization needs an LLM provider. The README does not specify a fixed provider, MCP client compatibility matrix, or a complete list of individual MCP tool names.

ZengLiangYi/ChatCrystal MCP server is intended for coding-conversation memory, not a general-purpose document or enterprise knowledge system. Imports depend on the supported source tools and their local data directories. Docker deployments that are reachable from other devices should use HTTPS through a reverse proxy for safer token transport.