An AI agent with its own computer.

Give it a task in a chat. It opens its own browser, signs in with logins you saved and gets it done. It asks you only for new sites, money and things only you can do.

Open source (MIT) ยท Runs on your own Claude plan

The Quokka Bot app: an agent called Errands answers when the Rijksmuseum is open on Sunday and what tickets cost, after the owner allowed rijksmuseum.nl for the task.
The agent's own browser, seen live: the museum's page with the prices it read.
Its computer, live

How a task goes

  1. You ask.

    Write the task like a message: check a museum's opening hours, download last month's invoice, compare three reviews.

  2. It works on its own computer.

    Chrome, a workspace and Claude Code run in sandboxes on your server. Watch it live from the app.

  3. It asks only when it must.

    A new site, a missing login or a payment shows up as a card in the chat. Answer with one tap, or type a message to skip it.

The safety rules sit outside the model.

The agent reads untrusted pages and holds your logins, so nothing depends on the model behaving.

01

Its own computer

Each agent gets isolated sandboxes for its browser, its shell and its CLI. Nothing runs on your Mac or phone.

The agent's own Chrome, seen live from the app, on the museum's gallery page.

02

Only sites you allow

All traffic goes through a proxy that blocks everything by default. The agent asks before it reaches a new site; you allow it for one task or for good.

A card in the chat: the agent asks to reach notebookcheck.net and theverge.com, with Allow for this task, Always allow and Refuse.

03

A vault the model can't read

Passwords and 2FA secrets are encrypted on your server. Quokka Bot types them into the page; the model never sees them.

A sign-in card in the chat with a secure form for a username, a password and an authenticator code.

04

Takeover

For a captcha, a text-message code or a card payment, take over the browser from the app, then hand it back.

The computer page while the owner has control: Hand back, and the Stop everything button.

05

Audit log and kill switch

Every action goes into a hash-chained log you can verify. One button stops the computer and cuts its network.

The activity log: the chain head hash with Verify chain, then entries for a takeover and for sites the owner allowed.

Runs on one server you control.

One Linux server runs everything: the app, the database, the proxy and every agent's computer. A server with 4 vCPU and 8 GB of RAM runs four agents. Tasks, files and logins are stored on that server.

Bring your own Claude plan. Each agent runs the official Claude Code CLI, signed in with your account.

Read the install guide

git clone https://github.com/quokka-bot/quokkabot.git /opt/quokkabot/srcexport QUOKKABOT_DOMAIN=bot.example.comexport QUOKKABOT_VIEW_DOMAIN=view.bot.example.com/opt/quokkabot/src/deploy/install.sh --setup-volume /dev/sdb

Use it from your Mac, your phone or a browser.

iPhone app Coming soon

Answer cards and take over from your phone.

Web app

The same app at your server's address, in any browser.

Open source, MIT licensed.

Read the code, report bugs and suggest ideas on GitHub.