HTTP API

The panel — 127.0.0.1:8080

Loopback only. No authentication: these endpoints start processes and write to disk.

method	path	purpose
GET	/	The control panel UI
GET	/api/status	Engine state, VRAM, installed models, catalogue, downloads
POST	/api/start?model=NAME&ctx=N	Stop any engine, start this model. ctx optional
POST	/api/stop	Stop the engine
POST	/api/download/{id}	Begin downloading a catalogue entry; returns immediately
POST	/api/download/{id}/cancel	Cancel; the part file is kept for resume
GET	/api/logs?tail=N	Last N engine log lines
GET	/api/logstream	Engine log as Server-Sent Events

start and stop take a lock and return 409 if one is already in progress. /api/status is safe to poll; the UI does so every two seconds.

The engine — 127.0.0.1:1919

llama-server’s own API. The two that matter:

• POST /v1/chat/completions — OpenAI-compatible.

• POST /v1/messages — Anthropic-compatible, including streaming and tool calls. This is what lets Claude Code connect without a proxy.

Also available: /health, /v1/models, /v1/completions, /v1/embeddings, /props, /slots, /metrics. Full documentation lives with llama.cpp.

The engine is started with a q8_0 KV cache and flash attention on, and where the build is new enough, the draft model’s own cache is quantised to q8_0 too — llama.cpp holds that one at f16 by default and does not inherit the main setting. See What actually decides whether a model fits for why. /api/status reports all three under engines.fixed, so the page states them once rather than on every row.