Point it at a model
OpenCLI ships with no model and no key. On a fresh install the first screen says so and offers the two routes below.
This is not a list of integrations — it is one protocol. Anything that answers
/v1/chat/completions and /v1/models the way OpenAI does can be pointed at,
and its models appear in the picker on their own. You do not declare them
one by one.
On this machine
The desktop app can install one for you: open Models, and it looks at what is already on the machine before offering to fetch anything.
By hand, with Ollama already running:
model = "qwen3-coder"
[model_providers.ollama]
name = "Ollama"
base_url = "http://localhost:11434/v1"
wire_api = "chat"
[[models]]
slug = "qwen3-coder"
model = "qwen3-coder:30b"
provider = "ollama"
context_window = 32768
Put that in ~/.opencli/config.toml and run opencli.
Nothing leaves the machine, and there is no key to keep.
A provider
Any of them works the same way. The pattern is a base URL and the name of an environment variable holding the key:
model = "my-model"
[model_providers.my-provider]
name = "My Provider"
base_url = "https://api.example.com/v1"
env_key = "MY_PROVIDER_API_KEY"
wire_api = "chat"
export MY_PROVIDER_API_KEY=...
opencli
In the desktop app the same thing is done in Settings, where the key is stored and never shown again.
Free allowances are the cheapest way to try a big model
Several providers give something away, which is how you run a model far larger than your machine could hold. What is free changes month to month, so this page quotes no amounts — check their own pricing pages.
The README carries the full list of providers, each one verified by asking it for its model list.
A model may not come from the company that made it
Xiaomi publishes MiMo, ByteDance publishes Seed, and neither runs a public OpenAI-compatible endpoint of its own. Both are carried by others — so you reach MiMo through OpenRouter, Novita, DeepInfra, PPIO or Hugging Face, not through Xiaomi.
Being reachable is not the same as being good at this
A provider that answers the OpenAI API but not its tool-calling part will chat and little else: this agent works by calling tools. Before committing to a model, check it — five minutes, one fixed task.