Rakazo review: self-hosting always-on bots with their own computers

Rakazo, which we make, takes a different shape from most agents on this list. Instead of one assistant on your machine, it runs several persistent bots, each with its own chats, memory, routines and a computer (browser, terminal and desktop) that lives in a container rather than on your host. It’s Apache-2.0, launched in August 2026, had 7,538 GitHub stars when we checked on November 8, and calls itself beta. We hold it to the same checklist as everything else here; this review is about running it yourself.

Who it suits

Rakazo fits people who want bots doing recurring work while they’re away: a bot that triages an inbox every morning, one that watches a repo, one that preps a weekly report. If you want a single chat companion on your laptop, it’s more system than you need. It is a web app with a database and a worker, not a single binary, and the setup reflects that.

Setup: one script, all Docker

The README install needs Docker Engine with the Compose plugin, curl and OpenSSL. No clone, no Node:

mkdir -p rakazo && cd rakazo &&
curl -fsSLO https://raw.githubusercontent.com/elie222/rakazo/main/infra/compose/install-images.sh &&
bash install-images.sh

The script downloads the Compose files, writes a .env with random secrets, and starts everything. You then open http://127.0.0.1:5173, create an account and connect a model. Rerunning it keeps your existing .env, which makes upgrades less nerve-wracking.

For bots that keep working when your laptop is shut, the same installer runs on a VPS: --prepare-only writes the config, you set your domain, and you put HTTPS in front of port 5173 (the production guide covers Caddy and host hardening). The desktop and mobile apps can then connect to your own instance. There is also a desktop app that runs the stack locally in Docker, and a Railway template.

Compared with OpenClaw or Hermes Agent, there are more moving parts (Postgres, a worker, the web app), but the installer hides most of them. The cost shows up later: there’s more to back up and more to monitor.

Isolation: the bot’s computer is the sandbox

On a self-hosted install, each bot’s computer is a local Docker container by default, with a virtual display and a full Chromium. The Docker computers doc sets ceilings out of the box: 2 GB of memory, 2 CPUs and 2,048 processes per computer, all adjustable. So the agent’s shell and browser never run on your host, unlike the defaults in OpenClaw or Hermes.

Two honest caveats from the same doc. First, a supervisor service holds the Docker socket to manage those containers, which is equivalent to control of the Docker host; it isn’t exposed on the host and needs its own token, but it is the component to protect. Second, the computers’ network isn’t locked down until you act: keeping them off your LAN and cloud metadata endpoints means running a host firewall script, which the default auto egress setting then picks up (or set SANDBOX_COMPUTER_EGRESS=restricted to refuse computers without it). On a cloud VM, do that step.

If you’d rather not run containers at all, computers can also run on a remote sandbox provider with an API key.

Approvals

Approvals are set in the app, not in a config file. Ask has the bot ask before anything that changes something, Auto uses a separate approval model to wave through on-task actions and ask about the rest, and Never ask runs everything. Approval requests appear as cards in the chat, and you can allow a tool or site permanently once you’ve seen it enough times.

Telemetry and privacy

A self-hosted install sends no product analytics or error reports unless the operator configures them; the desktop app checks for app updates. With a local model through Ollama or LM Studio, nothing leaves your machine except what the bot itself browses to.

Models and cost

You choose the model per bot: an existing Claude or ChatGPT subscription, an API key (Anthropic, OpenAI, OpenRouter, Grok), or local models. Self-hosting is free. The cost to watch is hardware: every active bot computer is a running desktop and browser, and each can use up to 2 GB of memory by default, so size the server for how many bots you expect to work at once.

Moving from OpenClaw or Hermes

Rakazo can import an agent from an ~/.openclaw or ~/.hermes folder or backup. You pick what comes across: instructions, memory, skills, MCP servers and scheduled jobs (which arrive paused). API keys are only read if you tick the box.

What’s still rough

It’s beta, with frequent releases, so expect change. Telegram, WhatsApp and iMessage are only available when you self-host, and its chat apps are a short list (Slack, Telegram, WhatsApp, iMessage, SMS and Feishu), against 29 channels for OpenClaw.

How it compares

For a single personal agent with the widest channel support, OpenClaw is simpler. For a container-per-agent design in a much smaller codebase, see NanoClaw. Agent Zero is the closest in spirit, giving its agent its own Docker Linux workspace. Our best self-hosted AI agent roundup puts them side by side.

Verdict

Rakazo is built for a team of bots that work on their own, with isolation on by default and no telemetry. It asks more of your server than a single-process agent, and it is still beta. If always-on, multi-bot work is what you want, it’s worth a look; if you want one assistant in your chats, start elsewhere.