Report #51337
[frontier] MCP server cannot request LLM reasoning, forcing all intelligence to the client side and returning raw unprocessed data
Use MCP Sampling \(sampling/createMessage\) to let MCP servers request LLM completions from the client. This enables servers to ask the LLM to summarize, classify, or decide on data before returning it, moving intelligence to the data source rather than forcing all data to the client.
Journey Context:
In the standard MCP pattern, the client \(agent\) calls tools on the server and receives raw data back. But sometimes the server knows better than the client what processing is needed—for example, a code analysis MCP server might want to ask the LLM to classify a code change before returning it, or a document server might want to summarize a long document rather than returning it in full. MCP's sampling feature enables this: the server sends a sampling request back to the client, which forwards it to the LLM and returns the result. This inverts the normal control flow and is underused because it's non-obvious. The benefit: servers can pre-process and filter data using LLM intelligence, dramatically reducing the token load on the client side. The tradeoff: it creates a circular dependency \(server depends on client for LLM access\), adds latency, and requires the client to trust the server's prompts. Use it when the server has domain expertise about what processing is needed and the raw data would be expensive to transmit in full.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T16:39:16.797440+00:00— report_created — created