Providers

Open Interpreter separates the provider from the model. The active provider decides where requests are sent, how credentials are attached, which wire API is used, and which bundled model metadata seeds the picker.

The source of truth is code, not this table:

• Built-in runtime providers: codex-rs/model-provider-info/src/lib.rs
• Generated hosted-provider catalog: codex-rs/model-provider-info/provider_catalog.json
• Hosted-provider generator inputs: codex-rs/scripts/write_provider_catalog.py and codex-rs/model-provider-info/provider_catalog_overrides.json
• Model metadata: codex-rs/codex-api/model_compatibility_catalog.json and codex-rs/models-manager/src/provider_catalog_models.rs

Built-In Runtime Providers

These providers exist even before the generated hosted-provider catalog is consulted.

OpenAI also has an onboarding preset named openai_api_key. That preset writes an OpenAI-compatible provider using OPENAI_API_KEY; it is a setup shortcut, not a separate built-in provider id in the runtime provider map.

Wire APIs

wire_api controls the HTTP request shape.

For custom providers:

[model_providers.example]
name = "Example"
base_url = "https://api.example.com/v1"
env_key = "EXAMPLE_API_KEY"
wire_api = "chat"

Use wire_api = "responses" for OpenAI Responses-compatible providers, wire_api = "chat" for OpenAI-compatible chat-completions providers, and wire_api = "messages" only for Anthropic Messages-compatible providers.

Authentication

Provider auth can come from:

• env_key, which reads a bearer token from an environment variable.
• experimental_bearer_token, when embedding a token is unavoidable.
• command-backed auth, which runs a local command and caches the returned bearer token.
• aws, only for amazon-bedrock.
• OpenAI auth manager state for providers that set requires_openai_auth.

Anthropic-style API-key auth uses the x-api-key header and automatically adds anthropic-version: 2023-06-01. Other API-key providers use Authorization: Bearer ....

Command-backed auth example:

[model_providers.example.auth]
command = "example-token"
args = ["print"]
timeout_ms = 5000
refresh_interval_ms = 300000

Generated Hosted Providers

The generated catalog is built from https://models.dev/api.json plus the live Moonshot model list at https://api.moonshot.ai/v1/models. The generator includes providers backed by supported AI SDK packages, excludes unsupported or local-only entries, requires a usable base URL, and keeps only models that advertise tool calling and text output.

The current bundled generated catalog contains these providers:

Model Picker Behavior

The picker uses provider metadata in two passes:

1. It asks the active provider's /models endpoint when possible.
2. It seeds or supplements results from the bundled provider catalog when the provider matches by Anthropic identity, base URL, provider name, or auth env var.

Generated model entries carry display name, description, context window, reasoning control, input modalities, and priority. The generator does not invent reasoning levels from a boolean reasoning flag; those controls come from compatibility metadata or explicit overrides.

Default Harness Selection

When no harness is explicitly configured, Open Interpreter may choose a native harness mode from the provider/model family:

You can override this with harness = "..." in config. See Harness for route compatibility.

Regenerating The Catalog

Run this from codex-rs when provider source data changes:

python3 scripts/write_provider_catalog.py
python3 scripts/write_model_compatibility_catalog.py

The Moonshot live source requires MOONSHOT_API_KEY or KIMI_API_KEY. Missing auth is a generator failure, not a reason to silently keep stale data.