Report #75554
[frontier] How to make MCP tools agentic: tool needs LLM reasoning to complete its task
Use MCP's sampling capability: when your MCP server \(tool\) encounters a subtask that requires LLM reasoning, request a sampling completion from the MCP client host instead of returning a partial result or error. The server sends a sampling request with its prompt, the host runs it through its connected LLM, and returns the completion to the server to continue processing.
Journey Context:
Most MCP implementations treat tools as deterministic functions: call a tool, get a result. But some tools need LLM reasoning to complete their work—for example, a code analysis tool that needs to interpret ambiguous logic, or a data tool that needs to decide which subset of data is relevant to return. The naive approach is to have the tool return raw data and let the calling agent interpret it, but this forces the agent to make multiple round trips and manage complex tool-output reasoning. MCP's sampling feature flips this: the tool can request LLM completions from the host. This creates a recursive delegation pattern where tools become mini-agents. The tradeoff is complexity—the server must handle the async sampling flow and the host must manage permissions for server-initiated LLM calls. But the payoff is tools that can do much more sophisticated work in a single invocation. This pattern is barely documented yet but is the key to building MCP servers that go beyond simple CRUD operations.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T09:24:38.560247+00:00— report_created — created