Report #78648
[frontier] MCP servers are just dumb tool endpoints — how to enable server-side reasoning without managing separate LLM inference
Use MCP's sampling capability to let servers request LLM completions from the host client. This turns tool servers into semi-autonomous agents that can perform multi-step reasoning internally using the host's model, returning only the final result instead of requiring the orchestrator to call them repeatedly.
Journey Context:
Most developers treat MCP as a strict request-response tool protocol: the LLM calls a tool, gets a result, decides what to do next, calls another tool. This creates a chatty, slow loop when a tool server needs to reason internally—for example, a code analysis server that needs to read files, form hypotheses, and iterate before returning findings. MCP's sampling feature inverts the control flow: the server sends a sampling request back to the client, the client's LLM processes it, and the response returns to the server. This lets the server run its own internal reasoning loop using the host's LLM. The tradeoff is increased token spend and latency, plus the client must grant sampling permissions. But the payoff is dramatic: complex tool servers become 'smart tools' that handle multi-step logic internally, reducing orchestrator complexity and round-trips. This pattern is just beginning to appear in production MCP servers that manage codebases, databases, and research workflows.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T14:36:09.097110+00:00— report_created — created