self-hosted AI agent

Keeps working after you close the terminal. Not require to open server port for web-UI. Use Destar Relay Web UI for quick up and running. Your own server, your own LLM, private & safe. One server for your team, private chats for each.

An AI agent that runs on your server. You talk to it from a browser. It reads files, runs commands, edits code — and asks before it changes anything.

bash

$ curl -fsSL https://destar.app/install.sh | sh

Debian · Ubuntu · any systemd box. Then open destar.app.

Windows PowerShell and Mac Terminal - coming soon!

youbrowser · destar.app

provided by destar

⇄

relay / direct connect web-UIour relay or direct to your server

agent in your serverdestar-run agent running in your server

Your LLM API glm-5.3 qwen3.8-max deepseek-v4-pro

sandbox user: destar-run, unprivileged, no sudo

read· edit· run

approval required

allow decline

// architecture

How it works

Install `destar-run`

curl -fsSL https://destar.app/install.sh | sh

Creates destar user + destar-run sandbox. Drops a systemd unit. No root after install.

Pair Destar Relay

The installer prints a 6-character code. Enter it at destar.app. The server connects to the relay over WebSocket — no inbound ports needed.

BYOL - Bring Your Own LLM

Add your LLM API key in settings. OpenAI, Anthropic, or any provider that speaks the chat completions format. The key stays on your server.

Invite your team optional

Add their email addresses in settings. Everyone works on the same server with the models you connected, each in their own chats. How sharing works

Sandbox boundaries

Process Runs as destar-run — unprivileged, no sudo, no other home dirs visible

Filesystem Read-only root. Writes confined to working directory

Resources cgroup v2 — CPU, memory and PID limits

Network nftables — internet allowed, loopback restricted to DNS + agent ports only

Secrets Sandbox has none. Keys live on the brain side, never passed to tool execution

// team

One server, your whole team

Set the server up once, then share it. Your team gets the same agent and the same models, and nobody reads anybody else's chats.

Invite By email address, from settings. They accept from the mail and the server appears in their account. No sign-up form: accepting creates the account

Chats Private to each person. The admin sees who is on the server and how much each has used, never what was said

Models Members pick from the models the admin connected. Keys and settings stay with the admin

Leaving Remove a member and their chats are deleted from the server at once. Members can also leave on their own

Trust Every member gets the whole agent in the same sandbox, so it reaches the same files for all of you. Share with people you work with

// clients

Three interfaces, one agent

Client	Platform	Status	Notes
Web	Browser	live	Full client at destar.app - via our relay
iOS	iPhone · iPad	soon	Push notifications for approvals. Native session UI
Android	Android 10+	soon	Push notifications. Same protocol, native shell

// install

Get started