![The llama.cpp panel for a busy chat model: the model name with its size and context, tiles for busy slots, token rates, requests waiting, tokens generated and draft acceptance, one card per slot with a bar for context in use, and the full metrics table.](/assets/devops-llama-panel.svg)

*Illustrative mock-up with made-up data, not a screenshot.*

## Features

- Loaded model at a glance: parameters, quantization, size, context length and slots
- Generation and prompt tokens per second, with lifetime averages
- Requests waiting and tokens generated
- Draft-token acceptance for speculative decoding
- One card per slot, showing how much of its context is in use

## For AI agents (MCP)

Connect Claude or another MCP client to Levirge DevOps and the agent can work with this plugin directly.

- `find_connection` finds the right connection by name or type.
- `llama_read` reads the server's health, properties, loaded models, slots and metrics.
- An agent can check which model is loaded, how busy the slots are and how fast it's generating before it sends work.

## Capabilities

- Reads the server's health, properties, models, slots and metrics.

## Limits

- Read-only: it can't run a completion.
- Model file paths and prompt text are never returned.
- The llama.cpp server must be reachable from the machine running devops-client.
- Load tiles need the server started with `--metrics`.
