The tools you never use

Claim: every tool, plugin, and integration you connect to an agent costs something on every session — context space, startup time, a longer menu the model has to read before it can begin — whether or not it is ever used. Left unmanaged, most of them are never used, and you pay for all of them anyway.

The incident

I counted the tool calls across a month of one project. Of 11,669 calls, 0.6% went to connected tool-servers, and only three servers were ever touched at all. Around fifteen more — a legal pack, a database integration, two source-host integrations for a repo that has no remote, several language servers for languages the project does not use — were loaded on every single session and called zero times. Each still added to startup, to sign-in noise, and to the list the model reads before starting work.

This runs exactly opposite to the direction that helped elsewhere: the model’s vendor removed most of the agent’s built-in preamble with no loss, because less to read helped. Meanwhile the project was loading a large menu, almost none of which it needed.

The rule

Treat the connected-tool list as something you curate per project, not a pile you accumulate. Turn off the packs this project has no use for. The test is empirical: look at what actually got called over real sessions — that count comes straight out of the session archive, the same record the other measurements here run on — and disconnect the rest. This is a different question than mining the archive for a model regression: same instrument, you are just asking what you loaded and never called.

Prior art

Verdict: PARTIAL — phenomenon and fix are KNOWN, even productized; the measurement and framing are the delta. Audit tools already prune by call count: unclog (2026) flags MCP servers uncalled in 30 days as “paying schema rent for nothing”; mcp-tidy does the same. The costs are measured: Scott Spence and EclipseSource (2025–26) report tool definitions eating tens of thousands of tokens before work starts, and RAG-MCP (arXiv:2505.03275) shows tool-selection accuracy collapsing with too many tools. The vendor-direction source is Anthropic’s context-engineering post (2026-07-24): “removed over 80% of Claude Code’s system prompt … with no measurable loss.” The delta: the specific measurement (0.6% of 11,669 calls used a server; ~15 loaded and never called) and the framing as the local inversion of that 80% trim — the same logic one layer down, at the tool-server layer.