Are you an LLM? Read llms.txt for a summary of the docs, or llms-full.txt for the full context.
Skip to content

Providers

Configure hosted model credentials, allowed models, and model selections in Amodal. These are platform settings; keep them out of amodal.json.

Supported Providers

The catalog is served by the platform and may include Anthropic, OpenAI, Google, DeepSeek, Groq, Mistral, xAI, and other configured providers. The exact models, prices, context windows, and credential sources shown in Amodal come from the platform catalog and org-level allowlists.

Configuration

Organization Providers

Open Organization > Model Providers to configure provider access:

  • Enable or disable each provider.
  • Choose a credential source, such as BYOK or Amodal-managed keys when available for the workspace plan.
  • Add API keys or provider base URLs where the provider requires them.
  • Select the models available to the organization.
  • Verify provider credentials before using them in runtime traffic.

Agent Model

Open the agent's Models page to choose a model, set its thinking budget and output-token limit, and bind models to the simple and advanced tiers. The page also shows provider verification, catalog metadata, and recent usage. These selections are stored by the platform.

If no agent-specific model is configured, the repository's platform default is Gemini 3.5 Flash (google / gemini-3.5-flash). Check the Models page for the selection used by your deployment.

Runtime Behavior

The hosted runtime receives the selected model from platform configuration when Amodal resolves a deploy. Local and programmatic runtime usage can still pass model overrides through runtime APIs, but starter agent repos should not include models in amodal.json.

Provider failover and routing policy are platform/runtime concerns. Use the Models page and Cost page to verify which provider/model handled traffic.

A catalog entry is not proof of working upstream access. A model can be listed while its provider rejects it as unavailable or unauthorized. Verify credentials and run an eval through the route the deployed agent uses.

For slow responses, inspect the session's model, token usage, and tool timing. External API calls and provider retries also contribute to latency.

Multi-Model Comparison

Use Amodal eval runs and experiments to compare providers. Run the same eval suite against each provider and compare quality, latency, and cost before changing production routing.

Tool Schema Compatibility

The runtime uses provider SDK adapters to translate messages, tool definitions and model replies. OpenAI, Anthropic and Google have separate adapters; other OpenAI-compatible providers use the OpenAI adapter with a configured base URL. A shared runtime interface does not mean that every provider accepts every JSON Schema feature.

Native OpenAPI connections send each visible operation's schema as a formal tool definition. Google middleware simplifies closed-tuple schemas in the model-facing copy because its SDK drops their positional item definitions. The runtime validates actual tool arguments against the original schema before execution. The middleware applies to the Google provider path, not one Gemini model ID.

Compatibility depends on the provider route as well as the model. A successful gateway test does not establish support for a direct-provider route or every OpenAPI schema. Use evals to check the route and schemas your agent uses, including write approval and denial.