# Sampling (client feature, 2025-11-25)

## Metadata

| Field | Value |
|---|---|
| Institute | Molodetz Institute - Research & Retrieval Division |
| Subject | mcp-client-support |
| Document type | Source Dossier |
| Tier | 1 |
| Source title | Sampling (MCP Specification v2025-11-25) |
| Source URL | https://modelcontextprotocol.io/specification/2025-11-25/client/sampling |
| Source publisher / author | Model Context Protocol authors (Anthropic-donated open project; site docs) |
| Source type | standard |
| Source date (published) | 2025-11-25 (version stamp; no other publish date stated) |
| Date researched | 2026-10-18 |
| Researched by | tai (agent), model: Claude Sonnet 5, on behalf of retoor |
| Research method | gather-subagent direct page fetch, parent synthesis from evidence pack |
| Confidence | High (partial read; findings cover retrieved sections only, see Cons) |
| Status | Active |

## Summary

Sampling lets servers request LLM completions through the client ("with no server API keys necessary") via the server-to-client `sampling/createMessage` request carrying messages, model preferences, system prompt, and max tokens. Clients must declare the `sampling` capability (plus `sampling.tools` for tool-using requests, which servers must not send otherwise). The `includeContext` values `thisServer`/`allServers` are soft-deprecated and servers should avoid them.

## Key findings

- Sampling is servers requesting "LLM sampling ('completions' or 'generations')… via clients… with no server API keys necessary."
- "Clients that support sampling MUST declare the `sampling` capability"; tool use in sampling needs `sampling.tools`, and "Servers MUST NOT send tool-enabled sampling requests to Clients that have not declared" it.
- Server-to-client request `sampling/createMessage` with `messages[]`, `modelPreferences`, `systemPrompt`, `maxTokens`; the client responds with `role`/`content`/`model`/`stopReason`.
- `includeContext` values `"thisServer"`/`"allServers"` are "soft-deprecated"; "Servers SHOULD avoid using these values."

## Pros / strengths

- Sensible trust design: the client holds the keys and the model policy, and the `sampling.tools` sub-capability gates the riskiest (tool-using) requests so servers cannot escalate unilaterally.
- Request/response shapes are explicit (messages, preferences, system prompt, max tokens in; role, content, model, stop reason out).
- Soft-deprecation of the broad `includeContext` values shows the spec actively narrowing context leakage.

## Cons / limitations / caveats

- Per the pack's Uncertainty notes this page was only partially read; the sampling-with-tools flow and deeper subsections were not read, so findings cover the retrieved sections only.
- No explicit author names or publish dates appear on the page; the date above is the version stamp in the URL.
- Deprecated in the 2026-07-28 revision along with roots and logging, and server-initiated requests are replaced by MRTR there; also, sampling needs a real LLM backend, which is far outside a minimal stdlib client's scope.

## Relevance to tai

Optional and LLM-backed, so safely omitted by a minimal client (no declared capability, `-32601` if asked). Notable for tai only as a future direction: a client with model access could one day offer sampling, but the deprecation in 2026-07-28 argues against investing here.

## Changelog

| Date | Change |
|---|---|
| 2026-10-18 | Initial dossier created. |