FutureOS is an open-source AI agent positioned as "one agent, everywhere you work," covering terminals, desktops, mobile devices, and chat applications. Its core is written in Rust, emphasizing local-first and open source.
Main capabilities include:
- Unified backend, multi-platform access: a single gRPC agent drives terminal UI, desktop applications, mobile applications, CLI, and IM bots, sharing the same sessions, memory, and skills across different devices.
- Execution boundary configuration: configurable file access approval and operating system sandboxing. The desktop defaults to Unrestricted (off); before handling untrusted content, it is recommended to choose Manual or Sandboxed. For the specific security model and sandbox limitations, see the official documentation.
- Long-running support: a built-in loop control plane provides persistent goals, event-sourced state, and verification gates, supporting tasks that run for more than 24 hours, such as starting a research task at night and viewing the results on your phone in the morning.
Usage:
- Installation: a one-line installation script is provided. On macOS you get an officially signed app (arm64/Intel); on Linux you get a .deb on Debian/Ubuntu (desktop app + unified future CLI); other platforms use a portable tar package; on Windows, a signed installer runs silently via PowerShell. After installation completes, future init runs. For detailed installation steps for each platform (desktop app, toolchain, GUI packaging), see the Build & Install guide.
- Configure models: the agent requires at least one model provider. Interactive setup supports FutureOS-hosted models or custom providers. FutureOS-hosted models use device flow login to automatically configure keys and the model list. Known providers can also be used by placing the API key under the provider name in ~/.future/agent/auth.json; for providers that require a user-specific base URL (such as Azure's YOUR_RESOURCE), add a baseUrl field in auth.json; for providers outside the built-in catalog, specify them fully in ~/.future/agent/models.json.
- Install skills (optional): skills are additional capability packs, loaded from ~/.future/agent/skills/. You can browse the catalog and install the skills you need, requiring only network access, no need to run an agent. Not installing it does not affect basic Q&A; skills merely add ready-made workflows. For uninstall/update and other commands, see the CLI reference.
- Startup and connection: The terminal and CLI client are lightweight gRPC clients that by default connect via per-user local IPC (a Unix domain socket on macOS/Linux, and a named pipe restricted to the current user on Windows). On Unix, FUTURE_AGENT_SOCKET is respected; on Linux, when set, $XDG_RUNTIME_DIR/future/agent.sock is used, otherwise it falls back to ~/.future/run/agent.sock (which is also the macOS default). An agent started with a separate FutureOS home directory (future agent --home DIR) is addressed through that home directory. The TUI and desktop app automatically start as a sidecar when no agent is running, and can also be started manually. Then start the terminal UI from the same future command.
- Remote/development mode: Passing --grpc-addr 127.0.0.1:50051 to future agent switches it to a regular TCP service, and clients connect via FUTURE_AGENT_GRPC_ADDR=127.0.0.1:50051. TCP is opt-in; never expose it beyond trusted interfaces.
- Unified entry point: future <cmd> is the unified entry point for every Rust component, including future agent, future tui, future channel, and future loop. Each command runs the same code as the standalone binary of the same name.
Other features:
- The TUI provides 40+ slash commands (see the full list via /help) and keyboard shortcuts.
- Released builds check for updates once in the background after the TUI starts; if a new version is available, a local notification and a link to the official installer are shown, and it is not installed automatically. Release builds check the release channel, while test/nightly builds check nightly. Local/source builds skip the check. Network failures are handled silently, future tui --offline disables the check, and non-interactive commands never perform the check.
Use cases: Suitable for users who need to seamlessly use the same AI agent across multiple devices (terminal, desktop, mobile, chat apps) for longrunning tasks, research, and development.
Limitations and notes:
- File access approval and sandboxing can be unrestricted by default; manual adjustment is needed when handling untrusted content.
- At least one model provider is required for it to work.
- Remote TCP mode should not be exposed to untrusted networks.
- For details such as the specific supported platforms, the list of model providers, and the skills directory, refer to the official documentation.
