Claude Code test: 587-token saving costs 4,225 in gateway setup

The test set out to check a design idea seen on X: put all of an organization's MCP servers behind one gateway, so the agent searches through it and enables only the needed tools. MCP here is a standard plug for connecting AI to external tools and data; each MCP server exposes a handful of callable tools. The author could not reach the implementation or scale described in that post from a sandbox, so instead built a small version locally with 40 tools — one real and 39 decoys with similar names — and measured token flow across four conditions.

The result inverted the author's expectation. Directly listing 40 tools added 587 tokens upfront (about 15.051 per tool), but that upfront figure covers names only, because descriptions are fetched just before a call. The gateway's list, by contrast, is a single JSON string carrying names, descriptions and schemas, which is why fetching it once added 4,225 tokens for 10,862 bytes across 40 entries. The extra fetch also added a turn, and the gateway's cumulative reused input reached about 182k versus about 139k for the direct method.

What did hold up was tool selection: with 39 look-alike decoys, the AI never called one, and even without any instruction on how to use the gateway, it reached the two-step list-then-invoke path in all three runs. The author is careful to note the limits — one sandbox, one server, 40 tools, one model, and three runs per condition — and that the list cost should shrink per call in jobs that call tools many times, a case not yet measured.