Models & providers

The active runtime calls the configured model provider. In embedded mode that is your process; in hosted or self-hosted mode it is the Agntz worker. Configure a provider by exporting its API key in embedded mode or saving it in Settings → Connections for a worker. Provider credentials are never sent by the hosted client on individual run requests.

Supported providers

This table was reviewed on 2026-08-17. The model picker queries each configured provider's catalog when one is available and falls back to this curated set if the catalog cannot be reached. OpenRouter's public catalog is loaded without a key. Perplexity exposes a fixed Sonar model enum rather than a list endpoint, and Azure requests use your deployment name, so those two remain configuration driven.

Picking a model in a manifest

provider is the provider id from the table. name must be the exact model id—or Azure deployment name—that provider expects.

model:
  provider: anthropic
  name: claude-sonnet-5
  maxTokens: 4096

OpenAI's gpt-5.6 alias resolves to gpt-5.6-sol. The Pro capability is a reasoning mode on GPT-5.6 rather than a separate model id. Use gpt-5.6-terra for a balanced default and gpt-5.6-luna for high-volume, cost-sensitive work.

Common model controls

These manifest fields are normalized through the AI SDK and forwarded when the selected model supports them:

model:
  provider: mistral
  name: mistral-small-2603
  temperature: 0.2
  maxTokens: 4096
  topP: 0.95
  stopSequences: ["<END>"]
  seed: 42
  maxRetries: 2

Unsupported settings may be ignored or reported as provider warnings. The normalized run result records the requested provider/model, the actual model reported by the provider, finish reason, response id, warnings, and usage.

Model-specific rules matter on the newest releases:

• Claude Sonnet 5 uses adaptive thinking and rejects non-default temperature, topP, or topK.
• Gemini 3.6 Flash and Gemini 3.5 Flash-Lite deprecate the sampling controls temperature, topP, and topK.
• Groq retired llama-3.1-8b-instant and llama-3.3-70b-versatile on 2026-08-16; use its GPT-OSS production routes.
• DeepSeek retired the legacy deepseek-chat and deepseek-reasoner aliases on 2026-07-24; use a DeepSeek V4 id.

Provider-scoped options

Use providerOptions for settings with no portable equivalent. Options are namespaced so switching the manifest provider cannot accidentally send OpenAI settings to Anthropic.

model:
  provider: openai
  name: gpt-5.6-terra
  providerOptions:
    openai:
      reasoningEffort: medium
      store: false

model:
  provider: anthropic
  name: claude-sonnet-5
  providerOptions:
    anthropic:
      thinking:
        type: adaptive
      effort: medium

The inner keys are passed to the selected AI SDK provider. Validate them against that provider's current documentation and keep them additive so a model switch can fall back to the common fields. Secret-like option keys are rejected; store credentials in Connections or Secrets, never in providerOptions.

The old model.options field remains accepted for compatibility but new manifests should use providerOptions.

OpenRouter — one key, hundreds of models

OpenRouter proxies commercial and open-source models behind a single API key. Set the key and reference a current model by its <author>/<model> slug:

export OPENROUTER_API_KEY=sk-or-...

model:
  provider: openrouter
  name: anthropic/claude-sonnet-5

model:
  provider: openrouter
  name: google/gemini-3.6-flash

model:
  provider: openrouter
  name: deepseek/deepseek-v4-pro

OpenRouter reports the per-request USD cost on every response, so traces in the UI show actual spend instead of a static estimate.

Attribution

By default, requests through OpenRouter are attributed to your app with the headers HTTP-Referer: https://agntz.co and X-Title: agntz. Override them through the provider's stored config:

{ "referer": "https://your-app.com", "title": "Your App" }

Provider model references

• OpenAI models
• Anthropic model selection
• Gemini models
• OpenRouter model catalog
• Mistral models
• xAI models
• Groq models
• DeepSeek updates
• Perplexity Sonar models
• Cohere models
• Azure deployment endpoints

Other providers and custom endpoints

Every provider supports a baseUrl override in its stored config—useful for proxies and OpenAI-compatible gateways. For arbitrary providers not in the table, supply a custom modelProvider implementation to createRunner.