Report #91127
[frontier] MCP server returning raw data dumps that overflow the agent context window
Implement MCP Sampling to let your server request LLM completions from the client model. Use it to pre-filter, summarize, or classify data server-side before returning it, keeping responses context-appropriate.
Journey Context:
Most MCP implementations treat the server as a passive data provider and the client as the sole intelligence. This leads to servers returning massive raw datasets that consume the entire context window. The MCP Sampling capability inverts this: the server requests the client's LLM to complete a prompt, enabling server-side intelligence. For example, a database MCP server can ask the model to classify which rows are relevant before returning them. Developers avoid this because it seems circular \(server asking client for LLM help\) and adds latency, but the tradeoff is worth it: you return 10 relevant rows instead of 1000 raw ones. The alternative—hardcoding filtering logic in the server—is brittle and cannot adapt to novel queries.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T11:33:09.551421+00:00— report_created — created