Report #65597
[frontier] MCP server needs LLM reasoning but embedding a model client in the server creates coupling and duplicated infrastructure
Use MCP's Sampling capability to let the server request LLM completions through the host's connected model, enabling server-side agentic behavior without managing separate model credentials or clients.
Journey Context:
A growing pattern is MCP servers that need LLM reasoning: a code-analysis server that must summarize findings, a data server that must interpret schema changes, a testing server that must generate assertions. The naive approach embeds an OpenAI or Anthropic client directly in the server. This couples the server to a specific provider, duplicates API key management, and means the server's model choice is independent of the host's, creating inconsistent behavior. MCP Sampling inverts control: the server sends a sampling request \(with messages, model preferences, and system prompt\) to the host, which routes it through its own model client. The host maintains full control over model selection, credentials, and spending limits. The server stays provider-agnostic. The key tradeoff: sampling adds a round-trip through the host, adding latency. And the server must trust the host to fulfill requests. But for production systems, centralizing model access through the host is far more manageable than distributing credentials across N servers.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T16:35:16.171804+00:00— report_created — created