Tool definitions are charged to context on every single turn, and they are the part of an agent setup that grows quietly as you add servers. This puts a number on that and hands back a two-part index — a small eager list plus schemas fetched on demand — with a trace replay to confirm the model can still find the right tool afterwards. It is a paid hosted endpoint rather than something you run, so what you hand over is catalog metadata, not your data.
A remote MCP server aimed at a problem you only hit once several servers are connected at the same time: every tool definition rides along in the model's context on every turn, and nothing in the client tells you what that adds up to. It reads a catalog — a server URL, a server card, or a `tools/list` response — reports the cost, and hands back a slimmed version. The surface is read-only by design: no shell, no SQL, no deployment or deletion tools.
- `analyze_catalog_budget` measures eager catalog tokens and estimates what progressive loading would save
- `generate_progressive_index` produces an eager tool index plus an on-demand schema manifest
- `replay_trace_budget` checks whether representative traces still route correctly against the slim index
- `compare_client_loading` returns client-specific notes for Claude, Codex, Cursor, Gemini CLI and custom clients
- `monitor_catalog_drift` compares the current catalog budget against a prior baseline and a CI threshold
Nothing to install. It is a hosted Streamable HTTP server at `https://mcploadbudgeter.space/mcp`, registered as `space.mcploadbudgeter/mcploadbudgeter-mcp` at version 0.1.0. Access is paid: complete checkout, claim the one-time bearer token, and send it as `Authorization: Bearer <token>`. The token is shown once.