Rakazo is an open-source AI bot platform that appeared on GitHub's weekly trending list in week 43 of 2026 with 1,166 new stars in seven days, reaching roughly 5,300 stars in total. Its one-line pitch is "Open-source Grok Bot alternative. Choose your own model and sandbox." In plain terms, it lets you run persistent AI bots, each with its own memory, routines and a computer to work on, on infrastructure you control.
If you have followed the race to give consumers an always-on agent, the shape is familiar from Claude Duet, OpenAI Dots, Meta Muse and Grok Bot. Rakazo is the self-hosted version: nobody else holds the conversation history, and you decide which model and which sandbox sits behind it. It follows an earlier open-source attempt at the same idea, CopilotKit's OpenBot.
Rakazo at a glance
| Question | Answer |
|---|---|
| What is it? | A self-hosted platform for persistent AI bots and a personal assistant |
| License | Apache-2.0 |
| Status | Beta |
| Stars | About 5.3k stars, 947 forks, 29 open issues, 27 open pull requests |
| Where does it run? | Web, Electron desktop, iPhone and Android |
| Models | Bring your own through Pi, optional OpenRouter key, or a supported provider |
| Sandboxes | Docker (default locally), E2B, Daytona, CreateOS, Boat, Railway, or a trusted local computer |
| Integrations | Composio or Pipedream Connect, plus MCP, OpenAPI and Treg tool sources |
| Voice | Your own ElevenLabs, OpenAI, Cartesia or Fish Audio key |
| Install | Docker, Compose, curl, OpenSSL, then install-images.sh |
What can a Rakazo bot actually do?
Per the repository README, each bot is a persistent entity rather than a chat window.
Memory, routines and history. A bot keeps its own conversations, memory and scheduled routines. That is the feature that separates a "bot" from a stateless chat: you can ask it to check something every morning and it remembers what it told you yesterday.
Computers. Bots get browser, terminal, file and graphical desktop access. There are shared Team Computers and isolated Private computers, so a household or a small team can share one environment while personal bots keep their own. This is where the sandbox choice matters: the bot acts through the computer, so whatever the sandbox can reach, the bot can reach.
Delegation. A bot can hand work to peer bots or to short-lived subagents, which is the same pattern coding agents use to split a task.
Voice. Voice mode covers speaking replies, dictation and calling a bot. Rakazo does not bundle a voice engine: you supply your own key for ElevenLabs, OpenAI, Cartesia or Fish Audio, which keeps the cost visible but means another account to manage.
Apps and tools. Integrations arrive through Composio or Pipedream Connect, or through tool sources you add yourself via MCP or OpenAPI. If you are new to the protocol, our MCP explainer covers what an MCP server is and what it can reach.
How do I install Rakazo?
The quick start in the README needs Docker Engine, the Compose plugin, curl and OpenSSL.
# from the repository: review the installer, then run it
./install-images.sh
# then open the app
# http://127.0.0.1:5173
For a VPS, the same installer works, and the README recommends putting HTTPS in front of port 5173. Building from source needs Node.js 22.22.2 or newer on the 22 line, 24.x, or 26 and above (23.x and 25.x are unsupported), plus pnpm 9 and Docker. The desktop app is an Electron build that can install the local stack or connect to an existing server.
A sensible first test, borrowed from how we test other agent tools:
- Install locally with the default Docker sandbox and a cheap model.
- Create one bot and give it a harmless routine, such as summarizing a public RSS feed each morning.
- Check what the Docker computer can reach on your network before adding any integrations.
- Only then connect an app through Composio or Pipedream and watch which scopes it asks for.
Which sandbox should I pick?
A server rack beside a small building, standing for the self-hosted infrastructure behind an open-source Grok Bot alternative
Rakazo's headline promise is that you choose. The tradeoffs are practical.
| Option | Good for | Watch out for |
|---|---|---|
| Docker (default locally) | Free, fast, easy to inspect | Shares your machine; a container is not a full boundary |
| E2B or Daytona | Throwaway cloud environments | Another vendor sees the work; usage costs |
| Railway | Always-on server computers | Requires PRIVATE networking, a team or workspace token, and is metered |
| Trusted local computer | Acting on your actual desktop | Gives a model-driven bot your real files and logged-in apps |
If you want stronger confinement than a plain container, pair it with a policy layer such as NVIDIA's OpenShell, or look at the Firecracker-based sandbox Perplexity open-sourced. Rakazo does not claim to provide those guarantees itself, and the README does not describe a permission model for what a bot may do with its tools beyond the computer boundary.
What are the limits and gotchas?
The README is direct about several, and they are the part of the project to read before you invest time.
- Beta software. Expect rough edges and breaking changes.
- Data loss on teardown.
docker compose down -vdeletes all Postgres data, so back up before you tear down. - Runtime constraints. Node.js 23.x and 25.x are unsupported; the Linux AppImage needs FUSE 2; the local-computer setup detects Docker but not Podman.
- HTTPS is mandatory on public servers.
- Metered pieces. Railway computers and Treg are usage-metered, and operators who resell Treg in hosted products need a written agreement.
- Size limits. Workspace exports reject files over 64 MiB.
- Language. Some marketing pages and the Electron setup and menu are English-only.
There is also a security point the README does not settle: a bot with browser, terminal and a logged-in desktop is exposed to prompt injection from any page or message it reads. The Hugging Face agent manager guide and our roundup of agent security platforms cover the pattern: scope credentials narrowly, keep the sandbox away from personal accounts, and review what the bot did.
For a way to detect and block risky agent actions, see AgentBeam, the agent security platform from the explainx.ai team.
Rakazo vs the alternatives
| Rakazo | OpenBot | Hosted bots (Grok Bot, Muse, Duet) | |
|---|---|---|---|
| Hosting | Self-hosted | Self-hosted | Vendor cloud |
| Model | Your choice | Your choice | Vendor model |
| Memory and routines | Built in | Check the repo | Built in |
| Sandbox | Docker, E2B, Daytona, others | Per its README | Vendor managed |
| Cost | Your keys and hosts | Your keys and hosts | Subscription or bundled |
| Maturity | Beta | Earlier project | Product |
Details for OpenBot are in our earlier post and its repository. The hosted bots are covered in the comparison linked above; their privacy tradeoffs are the main reason some people choose self-hosting in the first place.
What are people likely to ask before installing it?
Do I need an API key? Yes, in practice. Rakazo does not ship a model. You connect your own credentials through Pi, set an OpenRouter key if you want one gateway to many models, or pick a supported provider during onboarding. Your bill is whatever those providers charge, and a bot that runs routines on a schedule spends tokens even when you are not talking to it, so set a spending cap at the provider before you create your first routine.
Does my data leave my machine? The conversation history, memory and routines live in your own Postgres database. But anything you send to a model provider, a voice provider or an integration broker such as Composio or Pipedream leaves your network, and a hosted sandbox such as E2B, Daytona or Railway sees whatever the bot does inside it. Self-hosted does not automatically mean private: it means you choose where each piece lives.
Can I run it with a local model? The README describes bringing your own credentials and does not promise a local-model path, so check the repository before assuming Ollama-style support. Our guide to running agents on a machine at home covers what local hardware can handle.
Who is it for? Tinkerers who want to see what a persistent bot with its own computer can do, small teams that want a shared agent without a vendor account, and developers who would otherwise wire together a chat UI, a scheduler, a sandbox and a tool broker themselves. It is not for anyone who needs a support contract or a security review today.
What should I watch in the repository? The open issues and pull requests are the honest roadmap. With 29 issues and 27 pull requests open on a project that gained over a thousand stars in a week, watch how quickly maintainers respond, whether security reports get handled, and whether releases are tagged, since pinning a version matters more than the star count when a bot can act on your accounts.
Is it worth trying?
If you want to understand what a persistent, always-on agent feels like without handing your history to a vendor, Rakazo is a low-cost experiment: Docker, a cheap model key and an afternoon. If you need reliability, audit trails or a hardened permission model, it is too early. The star count tells you the idea is popular, not that the code has been tested by the kind of adversarial users a bot with a desktop attracts.
Related reading
- Claude Duet vs OpenAI Dots vs Meta Muse vs Grok Bot
- CopilotKit OpenBot: open-source Grok Bot
- NVIDIA OpenShell: install, telemetry and developer reaction
- Perplexity open-sources its Firecracker agent sandbox
- What is MCP? The Model Context Protocol guide
- Hugging Face agent manager guide
- AI agent security platforms in 2026
Details reflect the Rakazo repository as of October 23, 2026. It is beta software, so versions, requirements and star counts change quickly.