Your own AI that gets real work done

It wrote itself.

All 1.17 million lines of DeTars — running on your own machine — were written by its own engine. An unreliable AI can't build a working copy of itself. That's why you can trust it with real work.

1,170,000lines · written by itself
~50frontier models · switch anytime
100%memory stays on your disk

· Every number here is countable in the codebase — not slide-deck rhetoric

Why now

Today's model is a brilliant but unreliable intern.

Dazzling in flashes — but it forgets, fakes completion, and crashes halfway. Nobody hands real work to something unreliable. DeTars turns it into a team you trust with the real thing.

Living with it

Toss it one line. It runs the rest for you.

1
Input from anywhere
Feishu / Slack / Telegram / WeChat — one line or one photo
2
It runs in the background
Hours, even all night — you don't babysit it
3
It comes back to you
It lives for you — not the other way around.
📷Pet Care
via Telegram
Snapped a photo of my dog and sent it over
editing · writing captions
✓ Funny short video + social captions ready
detars · pack: pet-care
One motif: input → runs in background → comes back
Why you can trust it
01

It thinks for itself

Give it a goal — ‘research three competitors this quarter, deliver a sourced report.’ It breaks the work down, spins up specialists, tracks their progress, and hands back a finished draft. The how is reasoned on the spot, not a hard-wired script.

orchestrator.loglive
plan goal → 3 subtasks
agent A · filings
agent B · news
agent C · cross-check
pause · await A/B → merge
Real: it forms teams, splits work, pauses and resumes — actually running, not faked.
02

It doesn't fall over

Real work runs for hours: flaky networks, rate limits, sleeping laptops. Most tools lose hours to a single hiccup. DeTars treats trouble as a fact — back off, retry, continue; reboot and it resumes. Let it run overnight; check results in the morning.

runtime.recovery03:14
429 rate limited → backoff
✓ recovered, resuming step 7
sleep checkpoint saved
✓ resumed from checkpoint
Real: errors are facts to route around, not instant death.
03

It can't fake it

AI's worst trick: swearing it's ‘done’ when it isn't; citing a source that never existed. DeTars makes ‘done’ a receipt it must honor — delivered or not, backed by evidence or not. Every citation is sealed; change one character and it's instantly void.

completion.receiptverified
deliverable ....... ✓ present
evidence chain .... ✓ complete
sha256: 9f4c…a1e
peer review ....... ✓ signed
citation altered → rejected
Real: ‘done’ is a verifiable receipt, not a feeling.
04

It's yours — and no one can take it

It lives on your own machine and learns you over time. But everything it remembers about you needs your nod first; that memory is a file on your disk, not an account on someone's server. No one can brick it with a notice the way Rewind or Humane did.

~/detars/memorylocal
./profile.mem
./projects/*.mem
new memory → awaiting your ok
✓ works offline · nothing uploaded
Real: memory and identity are files on your disk — sovereignty baked into the architecture.
05

It evolves itself

It hunts its own bugs into permanent regression coverage, reorganizes its own memory, and — when a task needs a capability it lacks — generates, installs, and manages its own tools and skills. Every version is harder to beat than the last.

evolution.logauto
regression +1 case sealed
memory reindexed
skill needed: pdf-extract
↳ generated · installed · ✓
every version harder to break
Real: reliability that compounds — and new abilities it grows on its own.
06

It doesn't burn your budget

Long unattended tasks usually mean huge bills. DeTars routes to keep the model's KV / prompt cache hot, so repeated context isn't paid for twice — far fewer tokens per task, and cheaper as model prices fall. Let it run all night without watching the meter.

routing.cachelive
context 184k tokens
↳ prompt cache · hit
billed → new tokens only
repeat context: not re-billed
Real: caching-aware routing — fewer tokens billed, no runaway bills.
0
lines of working code · written by its own engine

We don't just say it can do hard work — it produced itself.

An unreliable AI can't write 1.17M lines of working code. The product itself is the ultimate proof of what it can do — and the number is countable in the codebase, on the spot.

DeTars is your permanent third-party terminal.
Yours, not rented.

A private AI team you own — always on, commandable from any chat window, and one that comes back to you on its own.