Run a local model server

What it is

A local provider is a model server on your own machine, such as llama-server. Loomcoding declares each server once, in your profile, and any project can use it.

What it is for

• Using a model without sending prompts to a hosted service.
• Switching between local servers with /provider.

How

1. Start your model server on the port that its entry expects.
2. Start loom in your project. It probes each server in the background.
3. Run /provider and pick the server. Use /model to pick another model id.

Example

One [[provider]] entry in providers.toml, with placeholder values:

default = "llama"

[[provider]]
id = "llama"
name = "llama-server"
kind = "openai_compat"
base_url = "http://127.0.0.1:8081/v1"
model = "<model-id>"
context_window = 65536
api_key_env = "<VARIABLE_NAME>"   # optional: the name of a variable, never the key

To declare reasoning levels for a local model, add them to the project loomcoding.toml:

[provider]
efforts = ["low", "medium", "high"]

Servers in the registry

The first start seeds these three entries. Their model names are placeholders: set your own.

id	Server	Port	Key
llama	llama-server (llama.cpp)	8081	None
strata	Strata	8080	Variable STRATA_API_KEY
polystrata	PolyStrata	8082	None

The key is never written in the file. A server on a loopback address may omit api_key_env, and Loomcoding then sends a placeholder, because llama-server does not ask for a key. A remote address always needs the variable.

Server states

State	Meaning
up (N)	The server answered with N models.
down	Connection refused, timeout, or an answer that is not a model list.
sin acceso (clave)	The server answered 401 or 403: the key is missing or wrong.

A down server keeps its placeholder model, so you can select it before you start it. The probes repeat when you open /model, /settings or /provider, with a 5-second cooldown.

Commands

Command	Default	Requires restart	What it does
/provider [id]	Not applicable	No	Switches provider and model. Without an id, opens the list with the state of each server.
/model [id]	Not applicable	No	Changes the model of the current provider. It accepts any id, including one that is not listed.
/effort [level or default]	default	No	Changes the reasoning effort, limited to the levels the model declares.
/settings	Not applicable	No	Opens the full form: provider, URL, key variable, model and effort.

Settings and defaults

Name	Default	Requires restart
default in providers.toml	llama	Yes
[provider] efforts in loomcoding.toml	Not set: the catalog decides	Yes after a manual edit
[provider] context_window in loomcoding.toml	Not set: the catalog decides	Yes after a manual edit
[provider] request_timeout_secs in loomcoding.toml	120 seconds; the maximum is 3600	Yes after a manual edit

Result

Symptom	Cause and fix
/provider says Providers are still loading	The probe has not answered yet. Wait a moment and try again.
A server that is running shows down	The port or the /v1 path in base_url does not match the server. Compare them.
The state shows sin acceso (clave)	Define the variable named in api_key_env before you start loom.
The project still uses the old provider	Its loomcoding.toml wins over the registry. Change it with /provider.

For the technical details of the registry, see the local models document.

Next steps

• Send your first prompt
• Connect a cloud provider