The MCP server lists its tools on demand: from 171,665 input tokens a turn to about 4,800, with the same choices

What changed

• Three tools stand in for the platform tools: search_tools (free, ranked, sees exactly what the key can reach), describe_tool (the full documentation) and call_tool (every gate, hold, scope and credit check the direct call has).
• A platform tool is still callable by name in either mode, and a guessed name is answered with the closest real tools.
• Per key: auto (on demand past 100 tools or 64 KB), always full, or always on demand — switchable on the API & MCP page.

The A/B behind it

Twenty real tasks against the real catalogue, with tool execution mocked: on demand, Claude Sonnet 5 chose the right tool 20/20, GPT-6 Luna 20/20 and Claude Haiku 4.5 19/20 — as accurate as with the full list — at about 4,800 input tokens a turn instead of 171,665.

The in-app autonomous loop had the mirror-image problem: a six-platform run registers about 319 tools against a cap of 120, and the trim made 7 of 12 test targets unreachable. Trimmed tools are now reachable through the same search, and 11–12 of 12 targets were reached on every model tested.

Frequently asked questions

Why is my MCP server using so many tokens?

Every MCP host loads the whole tools/list into the model's context on each turn. A server exposing hundreds of tools can cost over 100,000 input tokens a turn before the model does anything; listing tools on demand (search, describe, call) avoids it.

Does on-demand tool discovery hurt tool choice?

Not in our A/B: on twenty real tasks, Claude Sonnet 5 and GPT-6 Luna chose the right tool 20/20 with on-demand discovery, the same as with the full list, and Claude Haiku 4.5 19/20.