mcp-context-cost

markitdown — context cost

98 tokens across 1 tools — lean (< 1K). Measured 2026-09-30 under methodology v1.0.

An Anthropic request carries 64 of those tokens as tool definitions. What Claude makes of them is not published for this server: its Claude count is missing, or was taken against a capture this measurement has since replaced.

   
server (self-reported) markitdown
status measured
tokenizer tiktoken / o200k_base
launch command uvx markitdown-mcp
isolation docker · ghcr.io/astral-sh/uv:python3.12-bookworm-slim · network bridge · linux/amd64 · network enabled for package fetch; clean FS, no host credentials
env vars supplied none
canonical SHA-256 eef53f600eee8a14332be832c38139fe2092e219c3429c033befa8c41adf054e
category vendor-official
source https://github.com/microsoft/markitdown

Where the tokens are

tool tokens share description input schema output schema
convert_to_markdown 96 98.0% 18 31 31

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.

Over time

date tokens tools release measured in change
2026-08-16 64 1 not recorded not recorded —
2026-08-18 64 1 not recorded docker no change
2026-08-19 64 1 not recorded docker no change
2026-09-04 64 1 1.8.1 docker no change
2026-09-05 64 1 1.8.1 docker no change
2026-09-09 64 1 1.8.1 docker no change
2026-09-30 98 1 not recorded docker +34

Some of these sweeps predate the isolation column, so the conditions they were measured under are not on record.

Full series: results/history.csv.

Re-derive it

npx -y mcp-context-cost verify results/markitdown/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.

Badge JSON · All servers · Leaderboard · Methodology