Use lcpp

lcpp is the agent that runs models on the deployment’s own llama.cpp engine. It needs no vendor account, no sign-in and no install, and starts from the launch dialog like any other kind.

How it works

What you do	What happens
Pick “Launch llama.cpp” in the launch dialog	The session runs on the deployment’s engine. A stopped engine starts on the first turn
Enter a llama-server URL under “Your own connection”	Sessions connect straight to that server instead of the deployment’s engine
Turn “Use llama.cpp” off	lcpp can no longer be started

Decide whether to use it

“Use llama.cpp” on the llama.cpp card in ⚙Settings → Agents is on / off for lcpp as a whole. It is on by default.

Off removes it from the launch paths and refuses launches from any other route
Sessions already running keep running when you turn it off

Connect to your own llama-server

To use a llama-server on your own network, enter it under “Your own connection”.

The llama.cpp card: "Use llama.cpp" is on, and "Your own connection" holds http://192.168.0.10:8080, reported reachable with the model it found

Enter the URL (http://192.168.0.10:8080, say) and, if needed, an API key
Press “Check connection”. The build, the window and the models appear

While it is set:

The launch dialog’s model list is that server’s
lcpp sessions connect straight to that server. They do not go through the Control Plane, so limits your tenant administrator set on engines do not apply
The “Chat” pill at the top of the screen reports that server too: its state (Connected / Not reachable / Checking) and the model it found (Icons, badges and menus)

Clear the field to go back to the deployment’s engine.

Behavior

The card’s “Behavior” picks the default model, the models to exclude and so on (the items every agent shares). It has no RTK setting.

Before your first session

Before your first session, read lcpp: what hardware measurement found. In short:

With the engine instance stopped, the first turn takes minutes
The context window has a lower bound worth choosing
Only a few model families are confirmed to work

Good to know

It runs as Managed only. It cannot start in Terminal (CLI)
Your agent instructions are added to the system prompt every turn. Running sessions pick them up from the next turn (Write down how you work, once)
MCP server tools are not loaded until needed. When their descriptions run past a few thousand characters, it finds and uses what it needs through three tools, tool_search, tool_describe and tool_call, instead of sending them all every turn. That keeps a small context window free
The history shows the names of the tools actually called. lcpp’s own tools (read, edit, bash and so on) are always sent
An administrator can pin either behaviour for the whole deployment

When it goes wrong

The first turn does not come back

The engine instance is starting. Check its state on the “Chat” pill at the top of the screen. It can take a few minutes.

“Check failed” appears

Check the URL, the port and the API key. The server must be reachable from the workspace.