One line

You need a terminal and python3 (3.9 or newer, which every Mac has). Copy this, paste it into Terminal, press Enter:

curl -fsSL https://gotcomped.com/comped.sh | sh

No account, no package manager, nothing installed. It downloads about 150 KB of standard-library Python into a temporary directory, runs it, and deletes the directory on the way out. Ten seconds, and the only thing left on your machine is ~/comped/. The script is sixty lines; read it before you paste it. Anything after sh -s -- goes to the run: … | sh -s -- plan=claude-pro-20.

It checks the download against its published checksum before running anything. That proves the download arrived whole, not that this site is honest; for that, read the source.

With node: npx

If you already have node, there is nothing to download by hand and no shell to worry about, which also makes this the one to try on Windows:

npx comped

The package is the same standard-library Python in an npm wrapper: no node dependencies, no install script, and bin/comped.js does nothing but find an interpreter and hand it the payload. You still need Python 3.9 or newer on your PATH. Arguments work the same: npx comped plan=claude-pro-20.

The other way in: as a rote Play

comped is also published as a Play on Modiqo's registry, and that is the careful way in: rote prints every parameter and every path the Play will touch and waits for your yes, and the whole package sits in public where you can read it before you run it. The trade is a free Modiqo account, because rote will not run anything until you are signed in.

curl -fsSL https://gotcomped.com/run.sh | sh

That line installs rote if you don't have it, signs you in, and runs the Play. About a minute the first time. To be asked at every step instead, use the registry's own installer:

curl -fsSL "https://play.modiqo.ai/install?play=rajkaria/comped" | sh

Same code every way. The Play, the download and the npm package carry a byte-identical copy of the core, and CI fails if they ever drift apart. Same parameters, same card, same row on the board.

Or ask the agent you already have

If you spend your day in Claude Code, Codex, Cursor or anything else with a shell, you don't need the terminal line at all. Paste this at your agent:

Run comped on this machine: fetch https://gotcomped.com/llms.txt and follow it.

It reads llms.txt, a briefing written for agents rather than people, and takes it from there: it asks you first whether to put your score on the leaderboard, runs comped without needing an account, and reads the card back to you. That file also tells it what it must not do, including summarising your logs by eye and copying anything out of ~/comped into a message or a commit. It is short. Read it before you paste the line.

Try it on sample data first

If you'd rather see it work before pointing it at your own logs, sample logs travel with it: real in shape, made-up in content. This works through every door.

curl -fsSL https://gotcomped.com/comped.sh | sh -s -- \
  claude_dir=resources/fixtures/claude \
  codex_dir=resources/fixtures/codex \
  out_dir=comped-demo leaderboard=false

Or with node:

npx comped \
  claude_dir=resources/fixtures/claude \
  codex_dir=resources/fixtures/codex \
  out_dir=comped-demo leaderboard=false

Or, as a Play:

rote play run https://play.modiqo.ai/rajkaria/comped \
  claude_dir=resources/fixtures/claude \
  codex_dir=resources/fixtures/codex \
  out_dir=comped-demo

About two seconds, and a card with tiny numbers. Then the real thing:

curl -fsSL https://gotcomped.com/comped.sh | sh

Everything lands in ~/comped/. Run it again tomorrow and the card grows a line telling you what moved.

Reading your card

Top to bottom:

• The big number. What the last 30 days would have cost at the provider's public API prices. Not a bill: you're on a subscription and you paid what you paid.
• Your comp score, and your tier. That number divided by what your plan costs for the same window. 13× means your subscription paid for itself thirteen times over; the tier is the word for that.
• Which AI cost what, largest first, with a bar.
• Cache-read share. How much of what the AI read came from cache. High is normal and good.
• Active days and sessions, so a big number reads as either heavy use or a heavy week.
• Since last time, once there is a last time.
• Detected. Which AI you use and through which tools. None of it typed by you.
• If you're on… every plan your provider sells, scored at once, with the one it assumed marked. Find your row.
• Repeat offenders. Things you've asked your agent to do three or more times on different days, with what the repeats cost.

What it works out for you

You type nothing. Every request your AI tools log carries the name of the model that answered it, and that name says who made it. From that, comped knows whether you're on Claude, ChatGPT/Codex, Kimi, GLM, DeepSeek, Gemini, Grok, Qwen, MiniMax or Mistral, and which tools you read it through.

The one thing your logs don't record is which tier you pay for: a Pro session and a Max session look identical. So rather than ask, or peek at your account (it never will), the card prices every plan your provider sells and marks the most expensive one that fits as the safe assumption. Your real score is at least that. If you want the exact row on the headline, say so once:

rote play run https://play.modiqo.ai/rajkaria/comped plan=claude-pro-20

Say it once: a typed plan is remembered in ~/comped/comped-plan.txt and every later run uses it. Delete the file to go back to inferring. Paying for something the table doesn't list: a Kimi or GLM coding plan, a team seat? plan=usd:29 prices it.

Your tier

The score lands you in one of seven tiers, printed on the card and in comped-share.txt: Paying customer (under 1×), Break-even (1–2×), Comped (2–5×), Properly comped (5–12×), All-you-can-eat (12–30×), Hostage situation (30–80×), Please stop (80× and up).

The leaderboard

The last step of a run posts your score to https://gotcomped.com/leaderboard.html and prints your rank. The line in comped-share.txt is rewritten with it, so what you post already says where you stand.

Options you might change

Add any of these after the command as name=value. Everything has a sensible default.

The full parameter list, including where each tool's logs are read from, is on the developers page.

What it writes

All of it under ~/comped/ (or the out_dir you chose), and nowhere else. The card lists every path it wrote.

When it looks wrong

"no log directory found"

Expected if you don't use that tool: it's skipped and the run continues. If you do use it, its logs live somewhere unusual; point the right option at them: claude_dir, codex_dir, pi_dir, opencode_dir.

The number looks too low

Check days_back (it's 30 by default), then the "unpriced" list at the bottom of the report. A model that isn't in the price list contributes nothing to the total, on purpose: a guessed price would make the number worse, not better.

The score is lower than I expected

The headline assumes the most expensive plan that fits. Look at the If you're on… rows for your actual tier, or pass plan=.

No repeat offenders

The bar is deliberately high: three asks, two sessions, two days. Try repeat_threshold=2.

No PNG

The SVG card is always written and uploads to LinkedIn as-is. For a PNG, install rsvg-convert, or use a Mac where it's built in.

It asked me to sign in

That's rote's registry, so the Play can be fetched and verified. comped itself never signs in to anything; the one thing it sends is your score to the leaderboard, and only if leaderboard is left at true.

It says "not posted"

The card is done; only the leaderboard post failed, usually because the machine is offline or the post timed out. Run again when you're online, or leave it. The exact reply is in ~/comped/comped-rank.json.

I'm on the board as anon-xxxx

The run had no handle, which is what happens by default: nothing about you is assumed. Pass handle=yourname to claim a name. The rote path fills in your rote handle automatically. A re-run replaces your row, so the anonymous one goes away when a named one arrives from the same machine.

Privacy

• Reads your AI tools' session logs. Nothing else.
• Never reads ~/.claude.json, ~/.codex/auth.json or any credential, keychain or token file. Which AI you use comes from the logs; your plan is never looked up.
• Sends one thing. After the card is written, your score goes to the leaderboard: handle, comp score, tier, full-price total, plan and its price, detected providers and tools, days, sessions, cache share, and a random id that keys your row. No paths, prompts, model names or hostnames. It's saved to ~/comped/comped-rank.json before it goes. leaderboard=false makes the run entirely offline: no telemetry, no "anonymous usage", no version check, nothing.
• Writes only under the folder you choose, and tells you every path.
• Your messages are cut to 120 characters and hashed. Never the full text, never on a card.

How to verify each of those rather than believe them is on the developers page.