Rakazo is an open source, self-hosted AI agent app for persistent bots that keep their own memory, routines and history instead of forgetting everything when you close the tab. It is aimed at developers and self-hosting tinkerers who already hold a model API key and would rather run the agent layer, and the sandbox it acts in, on hardware they control. The same backend is reachable from a browser, an Electron desktop app, an iPhone app and an Android build. The project positions itself as an alternative to Grok Bot, Meta Muse, OpenAI Dots, Instinct, OpenClaw and Hermes Agent. The code is TypeScript under Apache 2.0, and the repository sits at roughly 5,400 stars and 950 forks.
What it does
The README's own feature list is the clearest statement of scope:
- persistent bots with their own conversations, memory, routines and history
- voice mode: spoken replies, dictation and calling a bot, using your own ElevenLabs, OpenAI, Cartesia or Fish Audio key
- shared Team Computers alongside isolated Private computers
- browser, terminal, file and graphical desktop access
- bots that delegate to peer bots or to short-lived subagents
- bring-your-own model credentials through a layer the README calls Pi
- app integrations through Composio or Pipedream Connect
Those are the authors' claims from their own README, not an independent evaluation. The channel does not run the tools it covers.
How it works
Two choices stay with you: the model and the sandbox. Model credentials are supplied by you rather than bundled, so the bot's reasoning comes from whatever provider you connect. The sandbox is the computer a bot works on: Docker on your own machine, or a hosted provider, and the video script names E2B, Daytona and Boat as options. Computers come in two shapes, Team Computers that several bots share and Private computers kept isolated.
Inside one of those computers a bot can drive a browser, run terminal commands, read and write files, and use a graphical desktop, which is what separates this from a chat window with tool calls bolted on. Bots can also hand work to peer bots or spin up short-lived subagents. All of the clients, web, desktop and mobile, talk to the stack you host.
Getting started
You need Docker, plus model credentials or an API key; a voice key is optional and only needed for voice mode. The project does not list hardware requirements, so there is no stated CPU, RAM or GPU floor to plan against, and the README does not name a local model runner even though it says the complete stack can run locally. Budget accordingly: the software is free under Apache 2.0, but the intelligence is not. You pay the model provider, and separately the voice provider and any hosted sandbox you choose instead of local Docker.
The repository is github.com/elie222/rakazo. The channel's summary describes a single setup script that brings the whole stack up in Docker, after which you open localhost, connect a model and start chatting. The saved README is cut before the install section, so take the exact command from the repository's own instructions rather than guessing at one. The mobile builds do not come from a clone: the iPhone app is on the App Store, and Android ships as an APK attached to the repository's releases, which means sideloading and allowing installs from unknown sources. There is a Discord for the community and a project site at rakazo.com.
When to use it / when not
The main catch is stated by the authors themselves: Rakazo is in beta. Combine that with a stack of browser, terminal, filesystem and desktop access, and you should treat early installs as something to keep off credentials and data you cannot afford to lose.
Use it instead of Grok Bot or a hosted assistant when the persistence matters and you want conversations, files and the sandbox to live on your own infrastructure. Don't use it if you only want a better chat box: a hosted assistant gives you that today without Docker, a key to manage, or beta rough edges.
Alternatives
The comparison set the project names is the current crop of hosted agent assistants: Grok Bot, Meta Muse, OpenAI Dots, Instinct, OpenClaw and Hermes Agent. Those arrive finished and managed, and you pay a vendor who also holds your data. Rakazo trades that convenience for control over the model, the sandbox and the storage. If the agent work you care about is specifically coding in a terminal, a dedicated coding agent is a narrower and more mature fit than a general teammate platform.
Take this one seriously if you are already paying for model API calls and have decided the agent runtime should be yours: self-hosted persistence, pluggable sandboxes and real browser, terminal and desktop access together are an unusual combination, and Apache 2.0 plus clients on four platforms means you can build on it. If you want something stable to hand to non-technical colleagues, wait until beta ends.