BACK_TO_FEEDAICRIER_2
llama.cpp WebUI adds native MCP client
OPEN_SOURCE ↗
REDDIT · REDDIT// 27d agoPRODUCT UPDATE

llama.cpp WebUI adds native MCP client

ANNOUNCEMENTPRODUCTGITHUB

llama.cpp's built-in `llama-server` WebUI now ships with a full MCP client, enabling local models to connect to external MCP servers and execute tool calls in an agentic loop. The feature, merged March 6 via PR #18655, also adds a `--webui-mcp-proxy` flag to handle browser CORS restrictions when connecting to remote MCP servers.

// ANALYSIS

Native MCP support in llama.cpp is a big deal — it brings the open-source local inference stack to parity with hosted AI tools that already lean heavily on MCP for agentic workflows.

  • –The WebUI can now connect to any MCP server, browse its resources, invoke its tools, and run multi-step agentic loops entirely on-device
  • –A `--webui-mcp-proxy` backend flag solves the browser CORS problem, a common pain point for web-based MCP clients
  • –Compatibility is model-dependent: models must support tool-calling schemas to participate; older or smaller models may silently skip the MCP context, which is already surfacing as a community pain point
  • –An active bug (issue #20475) means the CORS proxy breaks with `--api-key` set, causing 401 errors against auth-protected MCP servers
  • –A separate PR (#19373) to bring MCP support to `llama-cli` (the command-line interface) is still open, so the WebUI remains the only native entry point for now
// TAGS
llama.cppllmmcpopen-sourceagentinferenceself-hosted

DISCOVERED

27d ago

2026-03-15

PUBLISHED

27d ago

2026-03-15

RELEVANCE

8/ 10

AUTHOR

BeepBeeepBeep