Tool Schema Budget
Tool definitions are sent on every single request, before the user has said anything. Paste a tools array to see which ones dominate the payload and which fields are being re-described in tool after tool.
Output appears here as you type.
Runs entirely in your browser. Nothing you paste here is uploaded or stored.
Questions
Is what I paste sent anywhere?
No — and it is worth being concrete about why that matters here. A prompt, a captured response and a tool definition all contain the things people are most careful with: system instructions, customer data, internal endpoints. Every tool on this page is a pure function running in your tab, on a page whose Content-Security-Policy sets connect-src to ‘none’, so the browser will not let it open a network connection at all.
Why bytes and not tokens?
Because a token count computed here would be wrong. Token counts are model-specific and only exact from a count_tokens endpoint, and general-purpose tokenizers are off by enough to mislead — particularly on JSON, which is mostly punctuation and short identifiers. What you act on is the ranking and the share, and bytes give both exactly. Treat the numbers as relative, and get an absolute count from the API when you need one.
Why do repeated fields matter so much?
Because the definition is re-sent once per tool, on every request. An analysis of GitHub’s own MCP server found "owner" in 36 of 60 tool schemas and "repo" in 39 — the same few lines of JSON, paid for dozens of times. Shared fields are usually the largest avoidable cost in a tool set, and the easiest to spot once something counts them.
Is a long description a problem?
Usually not. The description is what makes a tool get called correctly, and trimming it to save bytes is a bad trade. It is flagged only so you can check the length is doing work rather than restating the schema in prose.
How much does this actually cost?
A report on the MCP spec repository measured about 1,000 tokens for a heavy tool and roughly 10,000 tokens for a 20-tool set before any user message, putting first-turn schema cost at around $390 across 2,600 conversations. The author’s point was that the context spent is capacity the model no longer has for reasoning — the bill is the smaller half of the problem.