540 tokens across 7 tools — lean (< 1K). Measured 2026-09-04 under methodology v1.0.
An Anthropic request carries 540 of those tokens as tool definitions, and Claude counts those at 1,267.
| server (self-reported) | example-servers/puppeteer v0.1.0 |
| status | measured |
| tokenizer | tiktoken / o200k_base |
| launch command | npx -y @modelcontextprotocol/server-puppeteer |
| isolation | docker · public.ecr.aws/docker/library/node:22-slim · network bridge · architecture not on record · network enabled for package fetch; clean FS, no host credent |
| env vars supplied | none |
| canonical SHA-256 | d9e46f0316917c1d140ffeebceefd4d31c16188780ac278baf347ff536015dd0 |
| category | official-reference |
| source | https://github.com/modelcontextprotocol/servers-archived |
| tool | tokens | share | description | input schema |
|---|---|---|---|---|
| puppeteer_screenshot | 142 | 26.3% | 11 | 117 |
| puppeteer_navigate | 129 | 23.9% | 4 | 111 |
| puppeteer_select | 65 | 12.0% | 9 | 43 |
| puppeteer_fill | 60 | 11.1% | 5 | 42 |
| puppeteer_evaluate | 48 | 8.9% | 7 | 27 |
| puppeteer_click | 47 | 8.7% | 6 | 28 |
| puppeteer_hover | 47 | 8.7% | 6 | 28 |
Each tool is tokenized on its own, so the parts do not sum exactly to the whole: the array adds its own brackets and commas, and the tokenizer merges tokens across object boundaries. The badge number is always the count of the whole array, never a sum of parts.
Measured 2026-09-14 against claude-opus-5 via Anthropic’s count_tokens (method tools-delta/v1).
| tokens | ||
|---|---|---|
| o200k, full capture | 540 | the badge number — every byte tools/list returned |
| o200k, Anthropic fields only | 540 | 0.0% of the capture is MCP-only metadata |
| Claude, same fields | 1,267 | 2.35× the badge number |
An Anthropic tool definition carries name, description, and input_schema and nothing else, so title, annotations, outputSchema, execution, and icons are dropped before the request — that is the second row. The third row is the same tools counted by Anthropic, which is larger than the second because Anthropic’s tokenizer is denser on this content than o200k_base and the API adds its own framing (at most 328 tokens of it fixed, measured against a single minimal tool). The two effects run in opposite directions, which is why the Claude number is not a fixed multiple of the badge.
| date | tokens | tools | release | measured in | change |
|---|---|---|---|---|---|
| 2026-08-16 | 540 | 7 | not recorded | not recorded | — |
| 2026-08-19 | 540 | 7 | not recorded | docker | no change |
| 2026-09-02 | 540 | 7 | not recorded | docker | no change |
| 2026-09-04 | 540 | 7 | 0.1.0 | docker | no change |
Some of these sweeps predate the
isolationcolumn, so the conditions they were measured under are not on record.
Full series: results/history.csv.
npx -y mcp-context-cost verify results/puppeteer/measurement.json
That re-tokenizes the published capture and checks the count and the hash. If it disagrees with the badge, the badge is wrong — open an issue and it gets corrected.