Desktop & Computer Use

The River Desktop App runs the same workspace you use in the browser — but with capabilities only a native app can provide. Local file access. Native notifications. A system tray. A global shortcut. And, most importantly, Computer Use: the AI can operate your machine on your behalf.

What the desktop app adds#

Computer Use#

Computer Use is the most powerful capability the desktop app unlocks. You ask the AI to do something that requires operating a real application — log into a vendor portal, fill in a form, scrape a page that doesn't have an API, post to a niche social site — and an agent on your local machine runs it for you.

How it works#

1. You ask SpaceChat or an Agent to do something that needs computer control.
2. The AI calls the computer_use_run tool, which creates a job in the cloud queue.
3. Your local desktop app polls the queue, claims the job, and runs it on your machine.
4. The agent takes a screenshot, sends it to Claude, gets back the next action (click here, type that), executes the action with nut.js (mouse + keyboard), and loops until done.
5. The result — what was accomplished, what was found, screenshots if relevant — is reported back as an entity in your workspace.

What it's good at#

• Tasks where there's no API, only a website
• Multi-step actions across native applications
• Logging into private portals to fetch data
• Filling forms with structured information from your Dataspace
• "Take this list of companies and find their LinkedIn pages" — even when the company list lives in your CRM and LinkedIn doesn't expose what you need

What it's not good at#

• Anything that has a clean API (use a normal integration)
• Tasks that take more than ~30 minutes (current cap)
• Actions on accounts that block automated browsers (recaptcha-heavy sites)

Permissions#

Computer Use needs OS-level permissions to inject mouse and keyboard events. On first use, River will prompt for:

• Accessibility access (macOS) — required for keyboard/mouse simulation
• Screen recording (macOS) — required for the AI to see what's on screen
• Input simulation (Windows / Linux) — handled per-OS

You can revoke these any time in System Settings. Without them, Computer Use is disabled and the rest of River works normally.

The Agent Panel#

The Agent Panel is a floating window in the system tray. Click the tray icon and it opens — 420×560, always-on-top, available from anywhere.

What's inside:

• Agents tab — quick-launch your most-used Agents
• Chat tab — a SpaceChat thread that runs against a local Ollama model (no internet required)
• Runs tab — live list of in-flight Jobs with progress and a Cancel button

The panel is what makes River feel like a coworker: you can hand off a task without context-switching to the full app.

Screen-aware AI#

Some Agents can be given permission to read your screen periodically — to understand the context you're working in across native apps. This is opt-in per Agent and per Space.

Patterns this enables:

• "What am I doing right now?" — surface relevant entities based on the app you're focused in.
• Cross-app capture — pull a quote from a Slack DM into a Doc with one shortcut.
• Active research — an Agent that watches the page you're reading and quietly builds a brief in the background.

Screen access is always granular. You decide which Agents have it, when they have it, and you can revoke at any time. Nothing leaves your machine unless an Agent explicitly sends it somewhere.

When you don't need the desktop app#

The browser version of River is feature-complete for most workflows. Skip the desktop app if:

• You work mostly inside the browser anyway
• You don't need Computer Use or screen access
• You don't drag files in/out frequently
• You don't want the system tray panel

You can always add it later — your workspace is the same one across browser and desktop.

Tips#

• Set the global shortcut to something you'll actually press. Cmd Shift J (Mac) or Ctrl Shift J (Win/Linux) is a sane default.
• The Agent Panel's local Ollama chat is great for sensitive prompts you don't want sent to a cloud model.
• For Computer Use, give the agent specific success criteria ("until the order confirmation email lands") rather than open-ended instructions ("buy stuff for me").