llama.cpp model manager

The llama.cpp extension is a hidden core extension that lets you manage local llama.cpp router models from the TUI. It is not listed in the bundled-extension manifest, but it is registered at boot as a built-in extension.

Core built-in

Shipped as part of the Selesai core runtime; not listed in the extension manifest.

Runtime surface: Slash command

Setup and prerequisites

Available only in interactive TUI mode.
Requires a running llama.cpp-compatible server.
Configure the server URL with /login llama.cpp before using /llama.
Hugging Face search and downloads need HF_TOKEN for gated models.

What it sets up

Registers the /llama slash command and the llama.cpp provider.
Reads provider auth from the model registry.
Restores the last successful server model catalog across restarts and refreshes it when the server is available.
Uses server-reported context limits for output capacity and records streamed token usage.
Fails gracefully when the server is unreachable, offering retry or close.

What you can configure

Core llama.cpp model management is configured via provider auth and slash commands. It is a hidden built-in extension, not a manifest entry.

Configuration locations and precedence

Provider auth
/login llama.cpp
Credentials and base URL are stored via the model registry's provider auth flow.

Settings

Key / path	Type, default, and values	Description
`baseUrl`	string Default: —	Llama.cpp server URL. /login stores the URL in provider auth env.LLAMA_BASE_URL. At runtime, LLAMA_BASE_URL env var (if set) takes precedence over that stored value.
`apiKey`🔒	string Default: —	Optional API key for the llama.cpp server. Stored in provider auth; falls back to LLAMA_API_KEY env var at runtime if no stored key exists.

Environment variables

Key / path	Description
`LLAMA_BASE_URL`	Highest-priority runtime override for the configured llama.cpp server URL. Also used as the default during /login if no URL is entered.
`LLAMA_API_KEY`🔒	Runtime fallback API key for the llama.cpp server.
`HF_TOKEN`🔒	Hugging Face token used to search and download gated models.

Command, tool, and shortcut controls

Key / path	Description
`/login llama.cpp`	Configure base URL and optional API key for the llama.cpp provider.
`/llama`	Open the TUI model manager. Only available in interactive mode.

Source evidence

src/core/llama/index.tsllama.cpp source
src/core/built-in-extensions.tsBuilt-in extensions
src/core/llama/client.ts

What you can do

List loaded, sleeping, and unloaded models on the configured server.
Load or unload models.
Download new models from Hugging Face, including gated models with HF_TOKEN.
Select quantization when a model offers multiple options.

Commands, tools, and shortcuts

/login llama.cpp — set the server base URL and optional API key.
/llama — open the model-management TUI. Not available outside interactive mode.
Provider auth is the primary configuration surface. LLAMA_BASE_URL overrides the stored base URL at runtime.

Limits and safety

Only interactive sessions can open /llama.
Downloading gated models requires accepting access terms at Hugging Face and a valid HF_TOKEN on the server.
Loading a model may unload other models depending on server capacity.
Local-server failures are reported as warnings, not crashes.

Source evidence

Selesai version: 0.10.0