unc 🦥

the uncle who's been here longer than the version control. writes less code, says less, ships more. now he rides shotgun in your agent.

unc is a loadout, not a tool. one command equips your AI agent with three things that already work, binds them under one always-on rule, and keeps an honest receipt of what it saved you.

curl -fsSL https://raw.githubusercontent.com/0ss/unc/main/install.sh | bash

restart your agent. done. works with anything that reads a rule and lets a hook log a turn — Claude, Codex, Cursor, Gemini.

the loadout

three tools that already work (unc installs them, never vendors them — they stay fresh) + one rule + one honest benchmark:

tool	cuts	by
ponytail	code you write	@DietrichGebert
caveman	output tokens	@JuliusBrussee
codedb	context in	@justrach

the rule (injected every turn, silent): ponytail + caveman + codedb, always. the benchmark: a Stop hook logs real output tokens + git churn per turn. no phone-home, the file is yours.

the receipt

unc stats --share

   🦥  unc wrapped   estimate

   280k   output tokens saved   ~$4.20

   output   ██████░░░░  -65% caveman
   code     █████░░░░░  -54% ponytail est
   context  ██████████  codedb 12 queries · ~18k not read

   34 turns · trust me bro · github.com/0ss/unc

Warm truecolor in a real terminal. The number is always derived from your real logged usage — an empty log shows nothing, and the card always wears its tier so an estimate can never masquerade as proof.

honest by design

unc will not lie to you, which is the whole point:

• You only do each task once, rule-on — so there's no true baseline to subtract. unc never invents one. (ponytail's own /ponytail-gain refuses to for the same reason.)
• ESTIMATE (what most people see forever): each tool's own published number on its own axis — caveman -65% output, ponytail -54% code, codedb leaner context — scaled to your real usage, labeled trust me bro. An extrapolation, not a measurement.
• MEASURED (if you ever run unc off a few times): words-per-line-of-code, rule-on vs rule-off, from your own logs. Median, normalized per line, confidence scales with sample size.
• Lines saved is always a ponytail lab estimate — you can't measure the lines you didn't write.
• $ is output-only. Cache reads are shown but never monetized (they're ~10× cheaper). Tune with UNC_PRICE.

Full formulas + sources: METHODOLOGY.md.

use it

unc is a shell command, not something you type at your agent:

unc stats          # numbers + honest savings
unc stats --share  # the receipt card
unc on | off       # toggle the rule
unc status         # one-liner, for a statusline
unc uninstall      # remove unc (keeps your log)

what it touches (and easy off)

unc is hooks + a CLI, not a skill you add inside Claude. The installer edits your global ~/.claude/settings.json to register three hooks, drops them in ~/.claude/hooks/, and puts unc on your PATH. That's it. Opting out is layered and easy:

unc off              # pause the rule (reversible; unc on to resume)
unc uninstall        # remove hooks + cli, keep your log
unc uninstall --purge  # also delete ~/.unc

dev

bash tests/test.sh

MIT. built on ponytail, caveman, codedb — go star those.