Inspection harness
Paste a tools/list. Get the metric table.
Everything runs in this browser tab. Your JSON is never uploaded, stored or sent anywhere. This page's security policy blocks all outgoing connections (connect-src 'none', form-action 'none').
No credentials. Paste a sanitized tools/list only. If the JSON looks like it contains a key or token, you'll see a warning. Remove it before you share any report.
1 · Input
Accepts a full JSON-RPC response {"result":{"tools":[...]}}, {"tools":[...]}, or a bare array [...].
Samples are real tools/list responses from the public npm packages @modelcontextprotocol/server-github 2025.4.8, server-filesystem 2026.8.31 and server-everything 2026.8.31, stored verbatim in /inspect/samples/.
2 · Options
3 · Results
No input yet.
| Stage | Exposed tools | Chars | Tokens o200k_base | Tokens cl100k_base | ~Tokens chars÷4 | Reduction vs raw | Cost |
|---|
4 · Markdown for your report
Run an analysis first.
How it's measured
- Raw: your input, parsed and written as compact JSON (
JSON.stringify), the same way the proxy measures the tools/list response line.{"tools":[...]}and bare arrays are first wrapped as{"jsonrpc":"2.0","id":1,"result":...}. - Compressed only:
rewriteToolsListRawwith compression on and tiering off (the proxy's--no-tier). - Tiered:
rewriteToolsListRawwith the proxy defaults: compress, then list the core tools plusprestige_expand_tools. Lists that already fit the core set are unchanged. - Algorithm: ported from prestigesf/prestige-mcp commit
5fa78ee45da8bdb95c585ab9240f17f73a4e5585(src/schema-compress.ts,src/tier.ts,src/intercept.ts) to/inspect/prestige-tier.js. Its output matched the TypeScript original byte for byte on 18,030 test cases, including the three samples (2026-10-02). - Tokens: tiktoken
o200k_baseandcl100k_basethrough js-tiktoken 1.0.21, served from this site. These are the same encodings as the prestige-mcp real-server audit.chars ÷ 4is the proxy's own rounded estimate. - Cost: tokens × calls ÷ 1,000,000 × price. You set the price. The default is only an example.