Run a whole office of AI agents
on your own computer.
They plan, research, build, check and review your work — passing it desk to desk while you watch. Free, local-first, and powered by the API key or CLI agent you already have.
Windows today · macOS and Linux next · no account, no sign-in, no telemetry
The floor, live
One sentence in. A finished file out.
The same relay your office runs — research, build, QA, review, your own coded checks, a revision, then the notes worth keeping.
A worked example, in the shape the app records: agent, step, verdict, tokens.
Audit a real job
Step through a job, one desk at a time.
A worked example of one job, step by step. Walk it forward, or switch the checks off and see what would have reached you.
Read up on what a booking page needs.
findings, not a page
What you get
An office, not a chat box.
Everything marked FREE works today, on your machine. The rest is being built in the open.
Every handoff, on screen
Agents walk to each other to pass work over. The relay records the same journey in numbers: verdicts, revisions and tokens per step.
QA and a reviewer argue first
Work is validated point by point, scored out of ten, and sent back for a revision when it falls short — automatically.
Files you can open
Pages, scripts, documents and PowerPoint decks land in the folder you choose. Reports save as Markdown or a printable HTML page, with the generated code inside.
Roles you describe or program
Invent a role in plain English, or hand it Python. Your code runs in its own process, so a mistake in it can never stop a job.
Memory and lessons
A knowledge keeper files what still matters in a month. A learning lab turns finished jobs into rules the office follows next time.
Pause, own, cap, cancel
Pause between steps and continue later. Choose who does a job, how hard to try, and cap what each agent may spend.
PowerPoint, actually written
Ask for a presentation and a real .pptx lands in your folder — not an essay about one, and not a script you have to run yourself. QA then opens the file and counts the slides, so a deck that is short or empty fails before it reaches you.
A cloud office, 24/7
Hand a job to your cloud office and it keeps going with your computer switched off. Results and the full report are waiting when you come back.
Your office on your phone
Send a job from the bus and read the result over chai — without opening a laptop.
Offices that talk
Each person keeps their own office and their own keys; the offices pass work between themselves. Nothing pooled in the middle.
Bring your own brain
Twelve CLI agents. Five kinds of key.
DAVE does not sell you tokens. It drives the agent you already pay for, or a model running on your own machine — and every desk can use a different one.
Or paste a key: Google Gemini · OpenAI · Anthropic · Ollama (free, on your PC) · any OpenAI-compatible endpoint. New model names appear as soon as your provider lists them.
Brains
Three ways to give your office a brain.
One key for the whole floor, a different brain per role, or a different one on every desk — mix them however you like, and change your mind any time.
One key, everyone
Paste a single API key and the whole office uses it. Most people start here and never move — a free Gemini key runs all six agents.
One brain per role
Give each department its own model: a strong one for building, a cheap fast one for QA and memory. The bill drops without the quality doing the same.
Every desk its own
Any agent can have its own provider, model, CLI agent or even its own key — your Claude Code subscription on the builder, a free key everywhere else.
Why people use it
Four things you will not find together anywhere else.
Plenty of tools run AI agents. These four are what make this one worth installing.
Watch it work, line by line
Who did what, what they passed on, what verdict came back, and what each step cost in tokens. Most tools hand you a wall of text and ask for trust. This one shows its working, and you can pause it mid-job.
The Learning Lab, after every job
When a task ends, one agent pulls out what worked and what failed and writes it down as a rule. Those lessons are fed into similar tasks later — and a lesson that leads to three failures is retired on its own. The office you use next month is not the one you installed.
Roles you can program
Twelve lines of Python, and your rule checks every build for ever — free, and with no tokens spent. It runs in its own process, so a mistake in your code is reported and skipped instead of stopping the job.
An honest "no" beats a confident guess
No key, no answer. A file cut off halfway fails instead of arriving broken. An image it cannot read is called unreadable, not described. Every one of these is covered by a test that runs before each release.
The plus points — and the limits
What it does well — and what it deliberately doesn't.
Most pages show you half the picture. Here is all of it, so you can tell in a minute whether DAVE fits how you work.
Where DAVE is strong
the reasons people install it and keep it
- You watch the work happen. Who did what, what they passed on, what it cost — no black box, no "trust me".
- Work is checked before you see it. QA validates point by point, a reviewer scores it, and weak work goes back for a revision automatically.
- It would rather fail than invent. No key, no answer. A half-written file is never handed over as finished.
- You choose the brains. One key for all, one per role, or one per desk — including the CLI agent you already pay for.
- Your rules, enforced forever. Write a role in English or in Python and it checks every build from then on.
- It learns your work. Notes and lessons from finished jobs change how the next ones are done.
- Costs stay visible. Tokens per step, per agent, per task, plus a spending cap on any desk.
- Nothing leaves your computer. No account, no telemetry; uninstalling is deleting one folder.
- Real files out. Pages, scripts, documents and PowerPoint decks you can open, plus a printable report of how they were made.
What it does not do — and why
every one of these is a choice or a date, not a surprise
- It runs while your PC is on. That is what keeps your work on your own machine. Want it round the clock? The Pro cloud office is being built for exactly that.
- Windows first, properly tested. One platform done well beats three done badly. macOS and Linux run from source today, and get installers next.
- It says "I can't read images" instead of guessing. A wrong description of your photo is worse than an honest no.
- It reads text, code, CSV and PDFs. Word and Excel are on the list; until then it tells you plainly, rather than half-reading them.
- Agents search, they do not roam. The researcher searches when you allow it. Nothing clicks around the internet on your behalf.
- Ten desks, three jobs at once. Sized for one person's real work, so your laptop stays usable while the office runs.
- You bring your own key. We sell no tokens, so we never mark them up — you pay your provider directly, at their price.
- The model you pick decides the quality. DAVE shows you exactly what each one produced and what it cost, so you can choose well.
- English interface for now. Every word is a plain string, so translations are welcome.
The first two are what Pro and the next builds are for. The rest are deliberate.
Private by design
Your work never passes through us.
Not a promise on a page — it is how the thing is built. There is no account, and today no server of ours sits in the path at all.
| What | Where it lives today | Does it reach us? |
|---|---|---|
| Your tasks, files and results | Your disk, in the app's data folder | No |
| Your API keys | Encrypted vault on your machine | No |
| Prompts and answers | Straight from your PC to the provider you chose | No |
| Agent memory and lessons | Plain files you can read or delete | No |
| Crash reports or analytics | Not collected | No |
The coming cloud office is opt-in and separate: you decide which job goes there, and the Classic office keeps working exactly as it does now. Read the privacy page →
Plans
Free today. Paid only for things that cost us money.
The office itself is free, for as long as it exists. No credit card, no trial clock, no feature held hostage.
Everything you have seen on this page, running on your own machine today.
- The whole office, up to 10 agents
- Roles you write in English or Python
- 12 CLI agents and 5 kinds of key
- Memory, lessons, reports, exports
- No account, no telemetry, no limits
Uses the AI key you already have. Read the 5-minute guide →
For people whose work should not stop when their laptop does.
- Cloud office, 24/7. Jobs keep running with your PC switched off
- Your office on your phone. Send work, read results, anywhere
- Wake up to finished work and a full report
- Inbox and calendar, kept on your own machine
- Everything in Classic, always included
No payment now. You will be told before anything costs money.
For a small team that wants each person's office to hand work to the next.
- Everything in Pro, per person
- Offices that hand work to each other
- Work passes between offices, never through ours
- Shared playbooks and roles
- Keys stay with each person, never pooled
Tell us the shape of your team; we size it with you.
You always pay your model provider directly — nothing on a free key or a local model. Every task shows exactly what it used.
Questions
Asked and answered.
Does my work ever leave my computer?
Only to the model provider you chose, and straight from your machine. We run no server in that path, and there is no account to sign into.
What does it cost to run?
The office is free. You pay whatever your provider charges — ₹0 on a free Gemini key or a local Ollama model, a few rupees a task on a paid one. Every task shows its tokens.
Do I need to know how to code?
No. Everything is buttons and plain English. Programming a role is optional, for people who want exact rules instead of judgement.
What happens if I close the lid mid-task?
The job stops where it is, and the office says so when you come back — it does not pretend the work finished. The agents live on your machine, which is the price of nothing leaving it.
What if an agent gets something wrong?
QA, the reviewer and your own roles check the work and send it back before you see it. You can pause a job in flight, or retry a failed one.
Windows only?
Windows is tested and released. The code is cross-platform and runs on macOS and Linux from source; packaged builds for both are next.
Open your office tonight.
Six specialists are already at their desks, waiting for a job. Setting up takes about five minutes, and the first task costs nothing on a free key.