AI NewsOpen sourceAnnouncement
Xiryn runs personal AI agents in Docker Linux desktops on your computer
Xiryn is a self-hosted take on the personal AI agent, with a Docker Linux desktop per agent, Telegram remote control, and an encrypted vault for the cards it spends.
Image: GitHub
Why it mattersA team can test the Muse and Grok Bot pattern on real tasks without putting staff logins or company cards into a vendor cloud, because the agent runs beside the laptop and sees only the fields it was approved for.
The personal AI agents every major lab shipped this year keep your logins and your credit cards on the vendor's servers. Xiryn is an open-source take on the same idea that stores everything on the user's own disk. The repository is at 180 stars on GitHub and names Meta Muse, xAI's Grok Bot and OpenAI Dots as the agents it replaces.
Each Xiryn agent gets its own Linux desktop in a Docker container, with a shell, a persistent home folder and a 1280 by 800 screen running Chromium. The agent works in that sandbox, so it never moves the user's mouse, reads their files or takes over the screen. The user can watch the agent's desktop live over noVNC in the browser, click Take over to pause it, type anything the agent got wrong and hand back with a note.
Where the private data sits
Conversations, memory, settings and keys live in one SQLite file on the user's disk. Saved logins, cards and personal details go into an encrypted vault, under AES-256-GCM, with the key held by the operating system's own secret store (macOS Keychain, Windows DPAPI, the Linux Secret Service). The repository says agents never see vault values. An agent at a checkout gets "Visa ···· 4242" and Xiryn itself types the card into the page.
Every payment is explicit. At a checkout the agent asks ("Atlas wants to pay ₹19,062 to MakeMyTrip") and the user picks Approve, I'll pay myself, or Decline, in the app or on Telegram. A saved login only fills on its own site, and an approved card only on the site it was approved for.
Telegram as the remote
The user pairs their own Telegram bot in Settings, which gives Xiryn remote control without opening a port or a public URL. The server long-polls Telegram from the laptop. When an agent needs a human (a login, a code, a payment), the bot sends a live link to the agent's desktop that works from the phone, and the user replies "done" when the step is finished so the agent carries on.
Model choice is open: OpenAI, Anthropic, Google, xAI, Groq, Mistral, DeepSeek, OpenRouter, Ollama, LM Studio and any OpenAI-compatible endpoint, picked per agent. The repository notes that running a local model through Ollama keeps even prompts off the network.
What setting it up costs
Xiryn is a Next.js server plus a SQLite database and runs on macOS, Linux and Windows. Setup is one script that installs Node, Docker and the agents' desktop image. There is no account, no sign-up, no telemetry, and nothing bills Xiryn's author: the user pays model providers directly for calls, so cost tracks real usage.
The README is honest about scope. The repository is new, the comparison table is based on public launch coverage, and the project asks for corrections. For a team deciding whether the Muse-style pattern is a fit for their work, Xiryn is a way to try the pattern on their own machines before putting a card and a session cookie on somebody else's servers.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
