# efaimo ai

> Know what you feed your agent. efaimo is an open-source CLI (Apache 2.0)
> that measures the context-window cost of MCP (Model Context Protocol)
> servers and Agent Skills, checks servers against the 2026-07-28 stateless
> MCP specification, and A/B tests whether a skill actually helps. v0.1.2 is
> live on npm (npx efaimo); the source is public at
> https://github.com/efaimo-ai/efaimo.

efaimo is pronounced eh-fai-mo and written 에파이모 in Korean. The brand and
this site are efaimo ai, also typed efaimoai, 에파이모 AI, 에파이모 에이아이,
or 에파이모에이아이; efaimo is the CLI the brand ships, and weigh, check
(MCP or skill), test and mcp are its commands.

Every number below is derived at build time, never typed. The weigh figures
come from one committed, machine-readable run against
`@modelcontextprotocol/server-everything` (2026-08-03): 13 tools
cost 1,120 o200k tokens serialized Claude-style
(1,142 as an OpenAI tools array, 1,607
as raw JSON), plus 312 tokens of server instructions,
measured separately and never summed with the rest. The grades are parsed out
of three further committed captures (check --mcp, check --skill, and check
--mcp against efaimo's own server), and the efaimo test deltas are quoted
from a live run rather than re-executed, with their provenance printed
beside them. (This paragraph said "one committed run" until 2026-08-02,
undercounting its own sources.)

## Pages

- [Home](https://efaimo.ai/): the thesis; one request replayed at true token
  proportions; a drafted schematic of where the instrument sits; a FAQ whose
  answers only restate the site
- [The cost](https://efaimo.ai/cost): the measured loadout in full: per-tool
  bars, description vs schema weight, and the 58-token markup gap
  accounted for exactly
- [The skills index](https://efaimo.ai/skills): every public Agent Skill in a
  pinned three-repository corpus graded by check --skill, 97% of
  36 scoring an A, a median of 1,632 instruction
  tokens loaded on trigger and 38.5 metadata tokens loaded at
  session start per installed skill; efaimo's own skills are graded by the same
  call in a separate table and are excluded from the corpus statistic
- [Commands](https://efaimo.ai/commands): five commands, four of them shown as
  real captured terminal runs; check --mcp grades quality and prints July 28 readiness separately
  as an ungraded migration diff, so the official reference server takes
  A (95) on quality with 4 items to migrate, efaimo's own
  SKILL.md grades A (100) from check --skill, efaimo's own MCP server
  takes A (100) with 3 migration items from its own grader,
  and efaimo test's first live deltas are +0 and +100 points
  (2026-07-17, claude-sonnet-5, 8 trials per arm,
  quoted from that run rather than re-executed; two scenarios is a
  demonstration, not a benchmark)
- [July 28, 2026](https://efaimo.ai/jul28): the major changes in the
  stateless spec, each row linked to the SEP or PR that made it
- [Method](https://efaimo.ai/method): reproduce, label, cite, dogfood; how a
  number is made and why unverifiable numbers never ship here
- [Privacy](https://efaimo.ai/privacy): two page counters (a cookieless
  Cloudflare beacon, and Google Analytics with ads signals off and no
  cookies in the EEA/UK), no cross-site tracking; the CLI sends no telemetry
  anywhere, and the only network calls it makes are to the target you name,
  and, when you ask for them, to the model APIs you configure with your own
  keys

## Source

- [GitHub repository](https://github.com/efaimo-ai/efaimo)
- [npm package](https://www.npmjs.com/package/efaimo)
- Contact: [hello@efaimo.ai](mailto:hello@efaimo.ai)