Description
You want an AI assistant that actually does the work, not one that only chats: it should run commands on your computer, drive a browser, operate phone apps that expose no API, and ask before it deletes, sends or pays. Meta's Muse takes the closed, cloud VM route. If you cannot get Muse or would rather not, try nanoMuse from a Zhejiang University team: one agent shared by your phone, your computer and your own server, open sourced under GPL-3.0.
The mobile app is built on the open source OpenMinis 1.13, it has no affiliation with Meta, and the little dragon mascot is its own. Anything that deletes, sends or pays raises a prompt, and you can allow it once, for the whole chat, or always for that kind of action. Passwords and codes stay yours to type, and at a login or CAPTCHA it stops and waits until you tap Done. Models can come from the relay's free allowance, your own key at 21 providers, or ChatGPT, Claude and Kimi plans. It runs on Android 8 or later, iPhone and iPad, macOS, Windows 10 or later and Linux, with a browser demo and Docker self hosting.
Does things: a built in Linux shell, browser, MCP and skills, plus Hands, which taps and types on a phone or computer screen for apps without an API
Asks first: confirmation before deleting, sending or paying, with approval scoped to once, per chat or always, and logins and CAPTCHAs handed to you
Across devices: speak on the phone and the task runs on the PC, or start a message with @Mac to send it there, with approvals returning to the device in your hand
Keeps going: goals checked on a schedule, routines that run while the app is closed, and a morning feed
Remembers you: memory is stored as Markdown files you can open and edit
Lives in chat apps: works from Feishu, DingTalk, WeCom and Telegram, so you can direct it from a conversation
A look of its own: an image model draws the character and a video model animates it, with a small dragon as the default
Self hosting: scripts/self-host.sh sets up your own server on one VPS in about an hour
The mobile app is built on the open source OpenMinis 1.13, it has no affiliation with Meta, and the little dragon mascot is its own. Anything that deletes, sends or pays raises a prompt, and you can allow it once, for the whole chat, or always for that kind of action. Passwords and codes stay yours to type, and at a login or CAPTCHA it stops and waits until you tap Done. Models can come from the relay's free allowance, your own key at 21 providers, or ChatGPT, Claude and Kimi plans. It runs on Android 8 or later, iPhone and iPad, macOS, Windows 10 or later and Linux, with a browser demo and Docker self hosting.
Features
Does things: a built in Linux shell, browser, MCP and skills, plus Hands, which taps and types on a phone or computer screen for apps without an API
Asks first: confirmation before deleting, sending or paying, with approval scoped to once, per chat or always, and logins and CAPTCHAs handed to you
Across devices: speak on the phone and the task runs on the PC, or start a message with @Mac to send it there, with approvals returning to the device in your hand
Keeps going: goals checked on a schedule, routines that run while the app is closed, and a morning feed
Remembers you: memory is stored as Markdown files you can open and edit
Lives in chat apps: works from Feishu, DingTalk, WeCom and Telegram, so you can direct it from a conversation
A look of its own: an image model draws the character and a video model animates it, with a small dragon as the default
Self hosting: scripts/self-host.sh sets up your own server on one VPS in about an hour
