Configure a model provider

A provider is the adapter between TAI and a model API. You need one provider and one model id before an agent can run. Start with the route that matches where your model lives; provider-specific tuning can wait.

Choose the shortest path

The dedicated plugins own vendor-specific behavior such as credentials, reasoning controls, prompt caching, and model discovery. If an API simply speaks the OpenAI wire format, the built-in adapter is usually enough.

Local and OpenAI-compatible models

Give the endpoint a meaningful id and declare its type. For Ollama:

# ~/.tailored-ai/config.yaml
providers:
  local:
    type: openai_compatible
    name: Ollama
    baseUrl: http://localhost:11434/v1
    defaultModel: qwen3

agent:
  defaultProvider: local

TAI does not start the model server or download the model. Make sure the server is already running and that baseUrl points to its /v1 API. An apiKey is optional and is only sent when configured.

The same shape works for a hosted OpenAI-compatible gateway:

providers:
  company_gateway:
    type: openai_compatible
    baseUrl: https://models.example.com/v1
    apiKey: ${MODEL_GATEWAY_API_KEY}
    defaultModel: team-default

agent:
  defaultProvider: company_gateway

You can define several endpoints under different ids. This is useful when a local model handles routine work and a hosted model handles harder requests.

Hosted providers

Install the adapter first. The installer downloads it into TAI's plugin directory and adds it to the plugins: list in your config:

tai plugin install @tailored-ai/provider-anthropic

Then add the provider's config and select it:

providers:
  anthropic:
    apiKey: ${ANTHROPIC_API_KEY}
    defaultModel: claude-haiku-4-5

agent:
  defaultProvider: anthropic

Keep secrets in environment variables or a .env file beside config.yaml; do not commit resolved keys. Each provider reference above lists its supported fields, credential behavior, and model-id rules.

Check the connection

Run one small request before configuring tools or automation:

tai -m "Reply with exactly: provider connected"

A completed response confirms that TAI could resolve the provider, authenticate, and invoke the selected model. If it fails, read the first error before changing anything: unknown provider ids usually mean the plugin is not installed; 401 or 403 responses point to credentials; connection errors point to baseUrl or the model server.

Use more than one provider

Set a provider and model for a specific agent when it should differ from the deployment default:

agents:
  researcher:
    provider: openrouter
    model: anthropic/claude-haiku-4.5

For automatic failover, give the default agent or a named agent an ordered models list. TAI tries the first entry, then moves down the list when a provider cannot be built or its request fails:

agents:
  default:
    models:
      - { provider: local, model: qwen3 }
      - { provider: anthropic, model: claude-haiku-4-5 }

See Configuration → Resilient model selection for the full precedence rules.

Your provider is not listed

First check whether its API is OpenAI-compatible. If it is, configure it inline with type: openai_compatible; you do not need a package for each vendor. If the protocol is genuinely different, build a provider plugin that implements TAI's small provider contract and registers a new id.