MCP tool schemas burn tokens on every call — new CLI makes the cost visible
Boris Pan explains that every MCP tool's name, description, and full JSON schema is re-sent to the model on every turn — not just at startup — and has built a CLI called `toolbudget` to make that hidden token cost visible.
Score breakdown
The post identifies a per-turn token cost that accumulates silently in every MCP server deployment, and the `toolbudget` CLI gives developers a concrete way to measure and manage it.
- 01Every MCP tool's name, description, and full JSON input schema is re-sent to the model on every call, not just at startup.
- 02Exposing 30 tools with verbose schemas can burn over 1,000 tokens on tool definitions alone before any useful work is done.
- 03A larger, noisier tool menu measurably degrades tool-selection accuracy by giving the model more near-identical options to confuse.
Boris Pan's post surfaces a hidden cost in MCP server development: tool definitions — each tool's name, description, and full JSON input schema — are re-injected into the model's context on every single turn, not once at startup. This means that an MCP server exposing 30 tools with verbose schemas can consume over a thousand tokens on tool definitions alone before the agent has performed any useful work.
Despite this, most developers add tools to an MCP server with no feedback on what it is costing in tokens or accuracy.
The problem compounds beyond raw token spend. Pan notes that a larger, noisier tool menu measurably degrades tool-selection accuracy, because the model faces more near-identical options and is more likely to confuse them. Despite this, most developers add tools to an MCP server with no feedback on what it is costing in tokens or accuracy.
To address this gap, Pan built a small CLI called `toolbudget`, published to npm, which makes the token bill from tool definitions visible to developers.
Key facts
- 01Every MCP tool's name, description, and full JSON input schema is re-sent to the model on every call, not just at startup.
- 02Exposing 30 tools with verbose schemas can burn over 1,000 tokens on tool definitions alone before any useful work is done.
- 03A larger, noisier tool menu measurably degrades tool-selection accuracy by giving the model more near-identical options to confuse.
- 04Most developers currently add tools to an MCP server with no feedback on the token cost.
- 05Boris Pan built a CLI called `toolbudget` to make the token cost of tool definitions visible.
- 06`toolbudget` is available on npm.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 11, 2026 · 08:34 UTC. How this works →