Custom LLM Providers

By default, OpenAI and Anthropic are pre-registered. The ProviderRegistry lets you add any Vercel AI SDK-compatible provider at runtime — no engine code changes required.

Registering a provider

import { createGroq } from '@ai-sdk/groq';
import { createProviderRegistry, configureProviderRegistry } from '@mcai/orchestrator';

// Start with built-in OpenAI + Anthropic
const providers = createProviderRegistry();

// Wire into the engine
configureProviderRegistry(providers);

Using providers in agent configs

Once registered, agents can reference the provider by name:

{
  "id": "fast-researcher",
  "model": "llama-3.3-70b-versatile",
  "provider": "groq",
  "system_prompt": "You are a research specialist...",
  "tools": [{ "type": "mcp", "server_id": "web-search" }]
}

Using provider options in agent configs

Once registered, agents can reference the provider by name:

{
  "id": "fast-researcher",
  "model": "claude-opus-4-20250514",
  "provider": "anthropic",
  "provider_options": {
    "effort": "max",
    "thinking": {
      "type": "enabled",
      "budgetTokens": 12000
    }
  },
  "system_prompt": "You are a research specialist...",
  "tools": [{ "type": "mcp", "server_id": "web-search" }]
}

Model inference and validation

Each provider declares a list of known models during registration. When an agent config omits the provider field, the engine infers it by exact match against these lists:

Model Name	Inferred Provider
gpt-4-turbo	openai
claude-sonnet-4-20250514	anthropic
llama-3.3-70b-versatile	groq

If no match is found, the default provider (anthropic) is used.

Note

The model list is advisory, not a strict allowlist. When resolveModel() is called with an unrecognised model ID, the engine logs a warning but still forwards the request to the provider. This allows using newly released models before the known list is updated.

Use addModel() to register new model names at runtime without re-registering the entire provider:

providers.addModel('openai', 'gpt-5');

Use supportsModel() to check if a model is in the known list:

providers.supportsModel('openai', 'gpt-4o'); // true

API reference

Export	Description
ProviderRegistry	Class — register, unregister, list, resolve models, validate
LanguageModelFactory	Type — (modelId: string) => LanguageModel
ProviderOptions	Type — { models: string[] }
createProviderRegistry()	Returns a registry with built-ins pre-registered
configureProviderRegistry(registry)	Wires a registry into the global agent factory

Next steps

• Agents — how agents use providers
• Adding MCP Tools — give agents external capabilities