Model Providers

Model providers are configured through the Models UI and stored as LLMConfig records.

Recommended Setup

Start with one provider and model that fits your team's quality, latency, and cost requirements, then add specialized configs only when needed. For a straightforward Codex setup, choose gpt-5.6-sol. When a GPT-5.6 model config does not specify a reasoning effort, the runtime uses medium.

Use higher reasoning effort for difficult implementation or review work when the additional latency is justified. Avoid assuming that a model name alone determines capability: runtime tools also depend on provider support, Chat mode, agent permissions, and configured integrations.

Claude Opus and Sonnet configs are alternatives for teams using Anthropic. Local Ollama and OpenAI-compatible configs are useful when data location, custom gateways, or local capacity matter.

Provider Matrix

OpenAI-Compatible Providers

OpenAI-compatible configs are for providers and gateways that expose an OpenAI-style Chat Completions API. The Models UI includes built-in presets for common hosted providers, inference gateways, and local runtimes, plus a custom option for anything else.

Use a preset when available so OpenVibely can fill the expected endpoint pattern and attempt model discovery. Use Custom OpenAI-Compatible when you need to supply your own base URL, model list URL, auth header name or prefix, extra headers, or extra request body fields. If discovery returns no usable model list, enter the model ID manually.

Public OpenAI-compatible endpoints should use HTTPS. Plain HTTP base URLs are accepted for localhost and private network hosts so local gateways can be used without public TLS.

Built-In Presets

OpenVibely ships presets for these OpenAI-compatible providers:

These presets are convenience defaults, not separate provider types. They all save as OpenAI-compatible model configs with preset-specific base URLs and discovery behavior.

OAuth Variables

Hosted OAuth can use provider-specific client IDs, client secrets, authorization URLs, token URLs, and scopes. APP_BASE_URL controls public callback origins, and OAUTH_REDIRECT_MODE controls hosted vs localhost/manual behavior.

Reasoning Effort

When a model supports reasoning modes, the model config UI shows an Effort dropdown. Models without reasoning support do not show the dropdown.

Codex (OpenAI) models

The default effort is model-dependent. GPT-5.6, GPT-5.5, and GPT-5.4 variants default to medium; older variants generally default to high. The environment variable OPENVIBELY_CODEX_REASONING_EFFORT sets a fallback effort when the model config does not specify one.

Anthropic (Claude) models

max is the Anthropic equivalent of xhigh.

If a configured effort level is not supported by the selected model, OpenVibely falls back to the nearest supported level automatically.