What

Tool responses from https://api.worldmonitor.app/mcp are large enough to consume a significant fraction of an LLM client's context window per call.

Why it matters

• Context budget: LLM clients (Claude, GPT, etc.) have finite context per conversation. A single tool call returning tens of KB of JSON pushes earlier turns out of context faster, and limits how many tools the model can chain in one task.
• Latency & bandwidth: large payloads add transfer time, especially on slower connections, and waste tokens on data the user didn't ask about.
• Cost: every byte returned is a token the user pays for, even when the actual answer is one number.

Repro

Call almost any worldmonitor tool (e.g. get_conflict_events, get_market_data, get_world_brief) with no filters. The response payload is many KB of JSON covering the full dataset.

Suggested directions

A few things that would help, individually or together:

1. Server-side filters via inputSchema (related to the other issue I'm opening): if get_conflict_events accepted country, since, limit, the LLM can ask for what it actually needs.
2. Sensible defaults: cap unconfigured calls at e.g. top-N most relevant items rather than dumping the full dataset.
3. Pagination: standard cursor / next_page shape so clients can ask for more on demand.
4. Summary mode: an optional summary: true parameter that returns counts and headlines instead of the full record set; the LLM can drill down with a follow-up call.
5. Field projection: an include/fields parameter so clients can fetch only the columns they care about.