Skip to main content
When a server calls ctx.sample(), the request lands on your client. Sampling adapters answer it with a real LLM in one line: 
import { Client, AnthropicSamplingAdapter } from '@prefecthq/fastmcp-ts/client'
import Anthropic from '@anthropic-ai/sdk'

const client = await Client.connect('http://localhost:3000', {
  handlers: {
    sampling: (params) => new AnthropicSamplingAdapter(new Anthropic()).handleSampling(params),
  },
})
This page is under construction. The outline below sketches what it will cover.
  • The adapters — AnthropicSamplingAdapter, OpenAISamplingAdapter, GoogleSamplingAdapter, and GenericSamplingAdapter for any custom completion function. Each translates MCP sampling params to its provider’s API, including tool-use round-trips.
  • Peer dependencies — provider SDKs are optional peers (install only what you use).
  • Options — modelSelector (static model name or a function of the server’s model preferences) and onToken streaming callbacks.
  • Capability advertisement — providing handlers.sampling makes the client advertise the sampling capability automatically.