SouravBeraAkaSpeed/xiryn

What it solves

Big personal agents like Meta's Muse, xAI's Grok Bot and OpenAI's Dots are closed, subscription-gated and run on vendor cloud machines — holding your conversations, logins and payment details. Xiryn offers the same idea as an open-source agent you run yourself: no account, no subscription, and nothing stored anywhere but your own disk.

How it works

One command installs Git, Node.js and Docker, builds a Linux desktop image for the agents, and starts Xiryn at http://127.0.0.1:6796. Each agent gets its own Docker container — a sandboxed Linux desktop with a shell, persistent files and a real Chromium browser — completely separate from your computer. A --local flag installs Ollama and downloads a model that fits your memory, so you can run everything free with nothing leaving your machine.

You bring any model: OpenAI, Anthropic, Google, xAI, Groq, Mistral, DeepSeek, OpenRouter, Ollama, LM Studio, or any OpenAI-compatible endpoint, and pick a per-agent model. Agents carry names, roles, standing orders and long-term memory stored in a single SQLite file. An encrypted vault (AES-256-GCM, key held by your operating system) stores logins and cards; agents never see the values, and every payment requires your explicit approval.

Routines run scheduled tasks like "every morning at 8, brief me on…" even with the browser closed. Telegram pairing via QR code lets you message agents from your phone — no public URL, tunnel or port forwarding. If you close the tab mid-task, the agent keeps going; you can take over its live desktop with one click, then hand back with a note.

Who it's for

People who want the convenience of a personal AI agent — building and debugging code, research with cited sources, writing, filling web forms, preparing bookings or orders — without giving their conversations, logins and payment details to a vendor's cloud. It also suits users who prefer their own API keys, fully local free models, or tinkerers who like disposable, self-hosted infrastructure.

Highlights

• Private by design: no account, no sign-up; conversations, memory, settings and keys live in one SQLite file on your disk. Run a local model and even prompts never leave.
• Never touches your computer: agents work on their own Docker desktop — no mouse moves, no screen takeover, no file access on your machine.
• Any model, your keys: 11 providers including local Ollama and LM Studio; free local models in one click with a progress bar.
• Disposable machines: reset an agent and its container is destroyed; auto-delete after every session while memory stays safe.
• Agents as a team: squads delegate to the right specialist, each on its own model and machine.
• Routines on a schedule: daily, weekdays, weekly, every few hours, or once — runs on its own and notifies you when done.
• Web research via Firecrawl with clean page reading and cited sources.
• Cost transparency: every reply shows its price; Usage page sums by day, model and agent; a monthly budget pauses paid models.
• Encrypted vault that is phishing-proof: saved logins only fill on their own site, approved cards only on the approved site; agents stop and hand you the screen at logins, payments and CAPTCHAs.
• Telegram remote control: live progress, screenshots and a working link to the agent's desktop from your phone.
• Voice input through your browser's speech recognition — on-device where supported, so your voice never leaves the computer.
• Free + pay-as-you-go APIs; MIT license; PRs welcome.