A← Index

deck

An open-source, self-hosted Grok Bot: my own AI bots, each with a computer of its own.

RoleSolo: design & build
TypeSelf-hosted AI bots, open source
StackRust · Axum · SQLite · React · QEMU/KVM
StartedNovember 2026

IWhat it is

Hosted bots such as Grok Bot feel smooth, but someone else owns the model, the data and the code. deck gives me the same kind of bot on my own server. It is open source under MIT and Apache 2.0. I am its first user, and anyone can host their own copy.

I talk to deck in a web app, on my laptop or my phone. I ask for work on the web: read a page, fill a form, log in, collect data or check something every day. I can watch the bot work, or leave and come back to the result. Text, files and screenshots arrive in the chat.

I run several bots, one for each kind of work: email, job hunting, and later marketing on LinkedIn or Instagram. Each bot has its own accounts, files, memory, instructions and permissions, so one bot never sees another bot's logins.

IIHow it fits together

One Rust service runs the agent loop for every bot, outside the bots' computers. Any OpenAI-compatible model endpoint can drive it. Each bot's computer is a virtual machine with a full desktop. It sleeps when idle and wakes when work arrives.

IIIWhat it is built to do

A smooth chat
The goal is a reply that starts in about 1 second, with progress on screen while the bot works. No dead waits and no silent failures.
A computer for each bot
Each bot keeps a persistent desktop. deck saves the computer to disk when the bot is idle and restores it when the next task arrives.
Owner control
Each bot has permissions, approval rules and a spending budget. A post, a payment or a sent message can wait for my approval. One inbox collects approvals, failures and results from all bots.
Built on bare
The only browser is bare, with no fallback. A bot reads pages as text first. When a site fails in bare, deck files a bare issue, so each failure makes the browser better.

A bot reports what it did and what failed. It never claims a result that it did not check.

IVWhere it stands

On 12 November 2026, the one-bot runtime passed its three stages: a full desktop with bare and saved state, a guest executor with durable commands and cancellation, and server integration with a queue, controls, idle save and wake. A strict run in a real virtual machine completed browser, shell and file work, saved the idle computer and restored it for the next task.

In one measured run, a working bot used about 531 MiB of VM memory, and the shared server used about 21 MiB. A restored task completed in about 3 seconds. I keep one bot lean first, so that many bots stay cheap to run. These are single observations on a small synthetic workload, not capacity numbers.

Still open: creating bots through a chat, real account integrations such as email, LinkedIn and Instagram, automatic spending accounting, memory, and load tests with many bots. The design covers them, but they are not shipped.

SignedArjun