feat(mcp): MCP Sampling — server-side LLM reasoning via connected client

[pgroene]

Summary

Enable Kyber's MCP server to use MCP Sampling to delegate LLM reasoning to the connected client (Claude, ChatGPT, Cursor), rather than always calling Kyber's own configured LLM API.

What is MCP Sampling?

The MCP Sampling primitive allows a server to ask the client to perform an LLM completion on its behalf. The client sends the request to its own LLM and returns the result — so the server gets AI reasoning without making an independent LLM API call.

Why it matters for Kyber

Scenario	Today	With Sampling
Summarize last night's motion events	Kyber calls its own LLM	Client LLM does it — no extra token cost to Kyber
Anomaly detection on sensor history	Requires Kyber-side AI config	Server pushes raw data, client reasons about it
Classify intent from ambiguous user input	Always through kyber_ask pipeline	Server asks client to pre-classify, then routes
Explain a complex automation	Kyber LLM prompt	Delegated to the already-connected Claude/GPT

Proposed implementation

• Add /sampling/createMessage\ handler to Kyber's MCP server
• Expose a \kyber_analyze\ capability that accepts raw HA context + a question and returns the client LLM's reasoning
• Use sampling for lightweight summarization tasks to reduce Kyber's own LLM call volume
• Add sampling request/response to the MCP call log debug UI

Acceptance criteria

• Kyber MCP server correctly sends \sampling/createMessage\ requests per the MCP spec
• Connected client (Claude Desktop tested) responds and result is used by Kyber
• Sampling calls appear in the MCP debug call log
• Feature is gated behind \enable_mcp\ flag / only active when MCP is enabled
• Works without sampling gracefully (clients that don't support it fall back silently)

Notes

• Neither HA's MCP nor Kyber currently supports this — implementing it would be a differentiator
• MCP SDK (\mcp\ package) already has sampling types; implementation is on the server handler side
• Sampling is human-in-the-loop by default in most clients (Claude asks permission) — document this behaviour

[github-actions]

🤖 Auto-triage

Field	Value
Type	✨ feature-request
Severity	🟠 high
Timeline	🗓️ long-term

This issue proposes a significant new functionality for Kyber to enable server-side LLM reasoning via connected clients using MCP Sampling.

_{Labels applied automatically. Feel free to correct them if the classification is off.}