Choosing a model

Pick a model with one line; the rest of your code does not change.

Hosted vendors

import { provider } from '@intellibrix/ai-sdk'

provider('anthropic:claude-sonnet-5')
provider('openai:gpt-4.1')
provider('google:gemini-2.5-pro')

Install the matching @ai-sdk/* package (@ai-sdk/anthropic, @ai-sdk/openai, @ai-sdk/google). They are optional peers, loaded lazily.

OpenAI-compatible endpoints

OpenRouter, Ollama, vLLM, LM Studio and many gateways speak the OpenAI protocol. Use compatibleProvider (needs @ai-sdk/openai-compatible):

import { compatibleProvider } from '@intellibrix/ai-sdk'

const openrouter = compatibleProvider('anthropic/claude-sonnet-5', {
  name: 'openrouter',
  baseURL: 'https://openrouter.ai/api/v1',
  apiKey: process.env.OPENROUTER_API_KEY,
})

const ollama = compatibleProvider('llama3.1', {
  name: 'ollama',
  baseURL: 'http://localhost:11434/v1',
  apiKey: 'ollama',
})

To wrap any other AI SDK language model, use aiSdkProvider(languageModel).

Models that write tool calls as text

Some open and free models do not use native tool calling. Instead they print the call in their reply, for example:

<|tool_call_start|>[read_file(path='a.txt')]<|tool_call_end|>
<tool_call>{"name": "read_file", "arguments": {"path": "a.txt"}}</tool_call>

Wrap the provider with textToolCalls and these become real tool calls that drive the loop:

import { Intelligence, textToolCalls } from 'intellibrix'

const intelligence = new Intelligence({
  provider: textToolCalls(openrouter),
  model: 'openrouter/free',
  tools,
})

Native tool calls pass through untouched; only steps without any are rewritten.
While streaming, markup is held back so users never see half a tool call.
Approvals and schema validation still apply to rewritten calls.
Pass parsers to add other formats, or knownToolsOnly: true to leave calls to unknown tools as plain text.