OpenAI-compatible endpoints

o4 can load custom OpenAI-compatible chat providers from ~/.o4/providers.toml. These providers use the /chat/completions protocol, but many vendors only support part of the OpenAI request shape. Use the optional compat block on each model to avoid sending fields that the provider rejects.

Example

[[providers]]
id = "xiaomi"
name = "Xiaomi MiMo Token Plan"
api_key_env = "MIMO_API_KEY"
base_url = "https://token-plan-sgp.xiaomimimo.com/v1"

[[providers.models]]
id = "MiMo-V2.5-Pro"
name = "MiMo V2.5 Pro"
reasoning = true
input = ["text"]
context_window = 258000
max_tokens = 8192

[providers.models.cost]
input = 0.0
output = 0.0

[providers.models.compat]
supported_parameters = ["stream", "max_tokens", "tools"]
supports_reasoning_content = false

[providers.models.compat.reasoning]
efforts = ["low", "high"]
ultra = false

Then run:

export MIMO_API_KEY="your-token-plan-key"
o4 -m xiaomi/MiMo-V2.5-Pro

Compatibility Fields

supported_parameters is an allowlist for optional request fields. When set, o4 only sends the listed fields from this set:

Parameter	Effect
stream / streaming	Send stream: true
max_tokens	Send max_tokens
temperature	Send temperature
tools	Send OpenAI-style function tools
tool_choice	Also enables sending tools
function_call	Also enables sending tools
reasoning_content	Preserve prior assistant reasoning content
reasoning	Also enables prior assistant reasoning content

You can also use explicit booleans:

[providers.models.compat]
supports_streaming = true
supports_max_tokens = true
supports_temperature = false
supports_tools = true
supports_reasoning_content = false

If no compatibility fields are configured, o4 keeps the legacy behavior and sends the usual OpenAI-compatible fields.

Reasoning Capabilities

reasoning = true says the model can reason, but it does not imply that the provider accepts an effort parameter. Declare the exact ordered subset only when the endpoint supports it:

[providers.models.compat.reasoning]
efforts = ["minimal", "low", "medium", "high", "xhigh", "max"]
ultra = false

Valid provider efforts are minimal, low, medium, high, xhigh, and max. ultra is separate because it enables o4 orchestration while sending the model’s declared max effort. When this metadata is absent, the settings UI offers only default and o4 omits the effort field instead of guessing from the model name.