Llama.cpp gets full MCP support for local agentic tools

Llama.cpp, the popular local inference engine for large language models, now fully supports the Model Context Protocol (MCP) across all protocols. The effort, spearheaded by contributor ngxson, tackles the trickiest part: stdio-based MCP servers. While HTTP servers were already accessible to the client, stdio servers require deep integration into the model's tooling pipeline. The team modified the llama-cli terminal client to route through the MCP server instead of a separate model serving path, then folded MCP support into the existing native tools server. The result, merged via pull request #26062, transforms llama.cpp's WebUI into a full-fledged agentic chat interface. Users can provide MCP server configurations either in a standard JSON config file or inline via command-line arguments for on-demand setups.

This opens the door to fully local, privacy-preserving AI agents. For example, connecting the Serena coding MCP server gives users a local-model-powered agentic coder that requires no external APIs or cloud dependencies. Developers can now build custom coding assistants, document analyzers, or task automators that run entirely on their own hardware. The integration leverages llama.cpp's native C++ performance, so even complex agent loops remain responsive. This breakthrough eliminates one of the last barriers to running autonomous agents locally: reliable, low-latency tool calling without cloud round-trips.