Llama.cpp adds native stdio MCP support to llama-server for local agent workflows
solyarisoftware · x · 2026-07-27
Llama.cpp adds native stdio MCP support to llama-server
The update turns llama-server into a lightweight agent harness for local GGUF models:
- It can launch MCP servers as local subprocesses.
- It discovers available tools and exposes them through /tools and chat completions.
- It reuses Claude/Cursor-style MCP configuration.
- It automatically restarts failed MCP processes.
The practical result is a simpler stack: local model + llama.cpp + local tools, without a cloud model, separate MCP bridge, or a large agent framework. The post also notes the model still needs strong tool selection, instruction following, and multi-step reasoning, citing Qwen3.6-27B or 35B as examples of models that may work well.
Related event: llama.cpp Adds Full MCP Support(2 posts)→
More from coding & agent
- OpenSpace targets a skill management layer for AI agents, not just more skills — solyarisoftware · 2026-07-27
- Can Chrome’s installed LLM drive a customer chat and call MCP tools? — DashinTheFields · 2026-07-27
- An unused gaming PC becomes a remote Claude Code workspace — rchardkovacs · 2026-07-27
- New agent-memory design moves retained state into the behavior-selection layer — nice2Bnice2 · 2026-07-27
- OMK open-sources a provider-neutral control plane for coding agents — DMAE1133 · 2026-07-27
- A vibe-coded Trevor Noah books page was rebuilt in Three.js with pure math — xiaohu · 2026-07-27