FutureOS: One AI Agent That Follows You Across Every Screen You Own
You've probably got an AI setup that works well in one place and falls apart everywhere else. Maybe you love your terminal agent but can't check on it from your phone. Maybe you've got a desktop app you like, but the moment you want to script something, you're stuck. FutureOS is an attempt to fix that fragmentation by running a single agent behind every interface you use.
What It Does
FutureOS is an open-source AI agent built around a single Rust core. That core exposes a gRPC backend, and everything else — the terminal UI, desktop app, Android and iOS apps, CLI, and IM bots — talks to the same agent. Same sessions, same memory, same skills, no matter which surface you're on. Start something in the terminal, pick it up on your phone.
The architecture is deliberately centralized. You configure a model once (from a catalog of 3,800+ across 140+ providers, or your own via models.json), define your skills and approval rules once, and that configuration applies everywhere. The Rust core handles the agent logic, the IM channel bridge, the loop control plane, the CLI, and the TUI.
Why It's Cool
The safety model is configurable, not assumed. This is the detail that stood out most. Agent tools like read, write, edit, and shell can be governed by approval rules and sandbox tiers — off, manual, or sandbox. The desktop defaults to Unrestricted, which the README is upfront about, and it tells you to switch to Manual or Sandboxed before working with untrusted content. Platform support is honest too: macOS gets Seatbelt, Linux gets Bubblewrap, Windows gets restricted-token write protection, and availability genuinely differs between them. That kind of specificity is rare and useful.
Long runs are treated as an engineering problem. Most agents are fine for a quick task and useless for anything that takes a day. FutureOS ships a loop control plane with durable goals, todos, gates, and monitors designed for runs of 24+ hours or more. It uses a deterministic should-run kernel, event-sourced state, and hard checks like evidence floors, acceptance contracts, and verify gates. There's lease liveness and multi-agent support. The pitch — kick off a research task at night, review results from your phone in the morning — is exactly the kind of workload that breaks naive agents.
Sessions fork like a repository. You can branch any conversation, clone it, and navigate the resulting tree over JSONL session history. If you've ever wished you could try two approaches from the same starting point without losing either, this is the feature for you.
The built-in skills cover real ground. Fifteen-plus skills ship out of the box: image reading and generation, PDF and Word parsing, web search, browser control, slides, software install, and the /future-loop orchestrator for long-run goals. That's a practical starting set rather than a demo.
It's one agent, not five. The "one backend, every surface" claim is the whole point. IM bots are included, which means your agent can live in the chat apps you already have open.
How to Try It
Installation is a one-liner with no source build required:
macOS and Linux use a single script that auto-detects the platform. On macOS you get the official signed app (arm64 or Intel); Linux gets its own build. Windows users should check the README for the current install path.
Once installed, the workflow is roughly:
- Pick a model from the built-in catalog, or point FutureOS at your own provider through
models.json. - Decide your approval posture. If you're working with anything untrusted, move off the default Unrestricted setting to Manual or Sandboxed.
- Start a session in whichever interface suits you — TUI, desktop, mobile, CLI, or an IM bot — and it'll be there in the others.
The README links out to deeper docs worth reading before you go far: the security model, the sandbox guide, the model catalog, and the loop control plane architecture. There's also a Chinese-language README if that's useful to you.
The project is at github.com/futuregene/future-os, licensed MIT plus Apache-2.0.
Final Thoughts
FutureOS is aimed at people who already use agents seriously and are tired of context living in one silo. If your work spans a terminal, a desktop, and a phone — and especially if you run tasks that take hours or days — the combination of a shared core, forkable sessions, and a real loop control plane is worth a look. The honest caveat is that safety defaults and platform guarantees vary, and the README says so, which means you should read the sandbox guide before pointing this at anything sensitive. For a project this early in its public life, that transparency is a good sign.
Follow @githubprojects for more developer tools and open source projects.