Config

Copy the example and edit it:

cp orouta.toml.example orouta.toml
orouta --config orouta.toml

--config defaults to orouta.toml in the working directory. Listen defaults are host = "0.0.0.0" and port = 11434.

host = "0.0.0.0"
port = 11434

[auth]
keys = ["sk-orouta-alice"]

[[upstream]]
id = "home"
base_url = "http://192.168.1.10:11434"
api_key = ""

[[upstream]]
id = "desk"
base_url = "http://192.168.1.20:11434"

There is no model list in the TOML. orouta calls GET /api/tags on each host and routes by those names. Put a given model on one host. Clients can send llama3 when the host reports llama3:latest.

To download a new model, POST /api/pull?host=<id> through orouta (see API) or run ollama pull on that Ollama machine. orouta will see it on the next tags refresh.

Each remote Ollama must listen on 0.0.0.0, not only loopback. Otherwise orouta gets connection refused on the LAN IP. How to set that on macOS, Linux, and Windows is in Expose Ollama. Localhost-only Ollama is fine for an upstream of http://127.0.0.1:11434.

auth.keys empty means no client auth. Otherwise send Authorization: Bearer or x-api-key.

orouta reloads the TOML about a second after you save it — no restart needed. Listen address changes need a restart. Hosts and keys changed in the UI live in orouta.overlay.json next to the TOML and survive reloads; see Manage.

Unknown names are 404. Paths with no model (/api/ps, /api/version, blobs) go to the first host in the file.

An upstream api_key is sent to that Ollama as Authorization: Bearer. The client’s key is stripped.

GET /api/tags and GET /v1/models are the union of what the hosts report.