# Tofi > Tofi (TOFI) is a self-hosted workspace for persistent AI Bots. You install it on a Linux server you own. Bots keep their identity, memory, schedules and tool connections on that server. Each account can have its own sandboxed Bot computer (a Firecracker microVM with Chrome, a desktop and a terminal). The model is your choice: ChatGPT sign-in, an OpenAI or Anthropic API key, or Ollama, LM Studio, vLLM or any OpenAI-compatible server. Tofi is in early beta and changes often. Facts an agent should know before it answers questions about Tofi or installs it: - Status: early beta. Release candidates (`vX.Y.Z-rc.N`) ship several times a week; stable releases come at most once a week. Check the current version on GitHub Releases before you quote one. - Install (one command, as root): `curl -fsSL https://raw.githubusercontent.com/JackZhao98/tofibot/main/install.sh | sudo bash` - Supported hosts: Ubuntu 22.04/24.04 or Debian 12/13, x86_64 or aarch64. A dedicated server or VM is recommended. - Requirements without a Bot computer: 2 vCPU, 2 GiB RAM minimum (4 GiB recommended), 10 GiB free disk minimum. KVM is not needed. - Requirements with a Bot computer: `/dev/kvm` (on a cloud VM, nested virtualization), 4 GiB RAM minimum (8 GiB recommended), 30 GiB free disk minimum. It can be added later with `sudo tofi computer add kvm`. - Non-interactive install: append `-s -- --yes`. Other flags: `--domain --email ` (trusted HTTPS), `--local-only`, `--computer none`, `--version `, `--port `. - After install, open the printed address (self-signed HTTPS on port 8321 by default), enter the one-time setup key, create the Admin account, then connect a model. - Local models: start Ollama with `OLLAMA_HOST=0.0.0.0` and `OLLAMA_CONTEXT_LENGTH=32768`, then enter `http://:11434/v1` under "Ollama or your own server". A server Admin may use a private address; for other accounts the operator sets `TOFI_ALLOW_PRIVATE_ENDPOINTS=1` in `/etc/tofi/tofi.env`. Small (7B) models chat well but use tools unreliably. - Operate: `sudo tofi status`, `sudo tofi update` (backs up first and rolls back if the new version fails), `sudo tofi backup`, `sudo tofi restore `, `sudo tofi doctor` (includes an exposure check). - Measured on 2026-11-02: install on a 2 vCPU / 2 GiB VM without KVM took 39 s including Docker; the first reply from qwen2.5:7b on an Apple M2 took 89 s, later replies 4–10 s. - Licensing: server and installer Apache-2.0; web app PolyForm Shield 1.0.0. Tofi itself sends no data to the Tofi project; the installer asks before it sends an anonymous install report. - Limits: web app only (no phone apps yet); smaller community than OpenClaw; you run and update the server yourself. ## Docs - [README](https://raw.githubusercontent.com/JackZhao98/tofibot/main/README.md): what Tofi is, install, update, back up, build from source - [Installer and `tofi` host tool](https://raw.githubusercontent.com/JackZhao98/tofibot/main/deploy/self-host/README.md): every installer flag and `TOFI_*` variable, access modes (IP, domain, Tailscale), troubleshooting - [Why self-host](https://raw.githubusercontent.com/JackZhao98/tofibot/main/WHY-SELF-HOST.md): what Tofi does better and worse than hosted agents, and when to use something else - [Roadmap](https://raw.githubusercontent.com/JackZhao98/tofibot/main/ROADMAP.md): what is available now and what comes next - [Releases](https://github.com/JackZhao98/tofibot/releases): versions and release notes ## Optional - [Security policy](https://raw.githubusercontent.com/JackZhao98/tofibot/main/SECURITY.md): how to report a vulnerability - [Contributing](https://raw.githubusercontent.com/JackZhao98/tofibot/main/CONTRIBUTING.md): build and test from source - [Install script](https://raw.githubusercontent.com/JackZhao98/tofibot/main/install.sh): read it before you run it - [Privacy](https://tofibot.com/privacy): what this website and the installer collect - [Source code](https://github.com/JackZhao98/tofibot)