The configuration file

~/.opencli/config.toml. Everything here has a default; the file only holds what you have changed.

The desktop app writes the same file, keeping your comments and formatting, so editing by hand and editing in Settings are the same act.

Providers and models

A provider is a base URL and the name of an environment variable holding the key:

[model_providers.my-provider]
name = "My Provider"
base_url = "https://api.example.com/v1"
env_key = "MY_PROVIDER_API_KEY"
wire_api = "chat"

Models declared under it appear in the picker:

model = "my-model"

[[models]]
slug = "my-model"
model = "provider-side-model-id"
provider = "my-provider"
display_name = "My Model"
context_window = 128000
reasoning_efforts = ["low", "medium", "high"]

slug is what you type; model is what the provider calls it. They differ often enough to be worth separating.

You do not have to declare them. Models are fetched from the provider's /v1/models. Declaring one is for pinning a context window, a display name, or the efforts it accepts.

Approvals and sandbox

approval_policy = "on-failure"   # untrusted | on-failure | never
sandbox_mode = "workspace-write" # read-only | workspace-write | danger-full-access

Explained in Sandbox and approvals.

Tool output

tool_output_token_limit = 10000

A budget for what a tool's output may add to the conversation.

Tokens are estimated, not counted. The tokenizer belongs to the model, and this runs models it has never seen. The estimate is four ASCII bytes to a token and one token per character otherwise — close for English and Chinese alike, erring towards over-counting elsewhere. Setting this to 500 does not cut at exactly 500 tokens; it cuts near it, in the same units the number is written in.

Seeing what is in effect

Settings shows every value that applies and where it came from, including the ones you have never set. Show raw config is the file itself.

The full reference

Every key this file accepts, with its default: Full configuration reference.