pmetal mcp
Start the PMetal MCP server for clients that speak the Model Context Protocol. The MCP surface exposes PMetal operations as tools for training, inference, datasets, quantization, merging, jobs, benchmarking, and serving.
pmetal mcpInstallation
Section titled “Installation”The MCP command is behind the optional mcp feature:
cargo install pmetal --features mcpTo combine it with the serving endpoint:
cargo install pmetal --features "serve,mcp"Tool Areas
Section titled “Tool Areas”- Device and memory inspection
- Model search, download, fit estimation, and local model listing
- Training, distillation, GRPO, RLKD, and embedding-training jobs
- Runtime job status, logs, learning-rate control, checkpointing, and graceful stop
- Dataset analyze/preview/validate/download/convert/filter/split/merge/sample/template/prepare
- Quantization, LoRA fuse, model merge, packed experts, Ollama exports
- Evaluation and benchmarking
- Serve startup and chat/generation workflows
See Also
Section titled “See Also”- pmetal serve — HTTP inference server
- pmetal train — Training jobs