Ollama

Rho can use models served by Ollama through its OpenAI-compatible API. Ollama runs locally by default. An API key is optional.

For direct hosted access with an API key or device key and no local server, see Ollama Cloud. If a local Ollama install is signed in with ollama signin, cloud models pulled as *-cloud still work through this local provider.

At a glance

Setting	Value
Provider	ollama
Auth	none, or optional ollama-api-key
Default API base	http://127.0.0.1:11434/v1
Model list	Installed models from native /api/tags (falls back to /v1/models)
Environment override	RHO_OLLAMA_API_KEY (optional key)

Setup

Install Ollama, start its server, and install a model that supports tool calls. For example:

ollama serve
ollama pull <tool-capable-model>

In Rho, run /login ollama. The first field is the API base, prefilled with the local default. Keep it or replace it with another host. The second field is an API key; leave it blank for a keyless local server, or to keep a stored or env-supplied key. Use /logout ollama to drop a stored key.

/login ollama

Then open /config and choose Refresh model lists. The model picker then shows the models returned by Ollama. Startup also refreshes that list when the cache is stale. You can also select one directly:

/model ollama/<tool-capable-model>

First-run setup does not write [providers.ollama]. The endpoint is stored only after this login, or when you add the table by hand.

Use another server

/login ollama is the usual way to set a custom API base. You can also edit ~/.rho/config.toml:

[providers.ollama]
base_url = "http://192.168.1.20:11434/v1"

Keep the /v1 suffix. Rho derives the native API root by removing that segment (http://host:11434/v1 → http://host:11434/api/tags and /api/show) and still posts agent turns to /v1/chat/completions. Bases that do not end in /v1 skip native discovery and keep the plain /v1/models list. /doctor still probes /v1/models. The URL must use http or https and cannot contain credentials, a query, or a fragment.

A stored key is sent as a Bearer token. Leave the key blank to keep a stored or env-supplied key, or when the host does not require one. Custom headers for secured remote endpoints are not supported.

Model metadata

Context window and thinking capability come from the local server. Complete /api/tags rows are enough; Rho calls /api/show only when tags omit context length or capabilities, reading model_info.<arch>.context_length when present. That value is the model's advertised maximum, not the server's current num_ctx. Models that advertise only embedding stay out of the picker.

Model compatibility

Ollama's OpenAI-compatible endpoint can serve many model types, but not every model supports the tool calls needed by a coding agent. Choose a model with reliable tool-call support. Rho does not download models or claim that every installed model supports tools, images, reasoning controls, parallel calls, or usage data.

Reasoning

Ollama accepts reasoning_effort values low, medium, high, max, and none. Rho sends the selected level when the model advertises a thinking capability. Shift+Tab and /config cycle it. When the model is missing from models.dev and the server does not advertise capabilities, Rho still sends the field so you can turn thinking off, mapping other Rho levels to the nearest accepted value. Rho omits the field for models whose metadata reports reasoning as not configurable. Omitting the field lets Ollama enable thinking on its own.

If Ollama omits optional usage data, Rho still handles the response. Run /doctor to distinguish a reachable server with installed models, a reachable server with no models, an unreachable server, and an invalid or unsuccessful response.