Unmute Bench

The benchmark that measures the human, not the model.

Every AI benchmark scores the model. Unmute Bench scores you: how much work you actually hand to Claude Code and Codex, from your own local logs.

curl -fsSL https://justunmute.me/unmute-bench.mjs | node --input-type=module -

One self-contained file (unmute-bench.mjs, no dependencies). Read it before you run it: it's the same code as src/, bundled.

  UNMUTE BENCH v0.1 · the benchmark that measures the human, not the model

  Median per active day           before       after
                                 Jul 11–Aug 9     Sep 26–Oct 25
  ────────────────────────────────────────────────────────────────
  SPD  agent sessions                  3  →       10   3.3×
  MPD  messages to agents             46  →      144   3.1×
  AHD  hours agents ran              4.9  →      9.4   1.9×
  JPH  agents steered per hour         1  →        3   3×
  PAR  agents running at once          1  →        2
  AD   active days                    21  →       29
  TTW  task types per week           3.4  →      5.4
  ────────────────────────────────────────────────────────────────
  Statistically significant (p < 0.001)

What it measures

Code	Metric
SPD	Agent sessions started per active day
MPD	Messages to agents per active day
AHD	Hours per active day with an agent running
JPH	Different agents you message per active hour
PAR	Agents running at the same time
AD	Active days in the window
TTW	Task types per week (code, research, comms, content, personal, other)

Medians per active day, after removing automated runs, repeats and bursts. Full definitions, cleaning rules and the significance test: METHODOLOGY.md.

Usage

curl -fsSL https://justunmute.me/unmute-bench.mjs -o unmute-bench.mjs

node unmute-bench.mjs                       # last 30 days vs the 30 before
node unmute-bench.mjs --split 2026-08-10    # 30 days before vs after a date (e.g. when you switched tools)
node unmute-bench.mjs --days 14             # shorter windows
node unmute-bench.mjs --card bench.png      # shareable black-and-white card (PNG needs Chrome; use .svg otherwise)
node unmute-bench.mjs --json                # machine-readable

If you use Unmute, the split date defaults to your first Unmute session, and Unmute's session index fills in history that Claude Code has already pruned.

Privacy

Runs entirely on your machine. Reads only your own messages (timestamp and the first 400 characters), never agent replies or files. Sends nothing anywhere. No dependencies at runtime.

Build from source

npm install && npm run build   # tsc → dist/
npm run bundle                 # esbuild → unmute-bench.mjs

Requirements

Node 18+. Codex archives (.jsonl.zst) are read only if the zstd binary is installed.

Submit your result

Run node unmute-bench.mjs --json and open a result issue with the output and your main setup (terminal, desktop app, Cursor, Unmute, other) to be added to the leaderboard. Only the numbers are needed, never your logs.

Cite

@misc{unmutebench2026,
  title  = {Unmute Bench: A Benchmark of Human-to-Agent Work},
  author = {Patel, Arpit},
  year   = {2026},
  note   = {Version 0.1},
  url    = {https://justunmute.me/bench}
}

License

MIT