Skip to main content
FlowDeck includes an AI agent that can drive the full Apple development toolchain from a conversation. Instead of clicking through menus or remembering commands, you describe what you want and the agent does it.

​
What the agent can do

Build & Run

Build your app, fix compiler errors, launch on a simulator or device, and stream runtime logs — all from a single prompt.

Testing

Run your full test suite or specific tests. The agent reads failures, traces them to the source, and fixes bugs in production code or tests.

UI Automation

The agent sees the simulator screen, taps buttons, types text, scrolls, and captures screenshots to verify results.

Simulators & Devices

Boot simulators, install and launch apps on physical devices, manage runtimes and simulator state.

​
How to talk to the agent

Open the chat panel and type naturally. The agent understands intent — you don’t need to know FlowDeck’s commands. Some prompts that work well:

​
Supported agents

FlowDeck drives an agent CLI you install and authenticate yourself. Five are supported: All five appear under Settings → AI → Agents, whether or not they are installed. An agent FlowDeck cannot find on this Mac is shown but switched off, and it keeps your choice — reinstall it, refresh, and it comes back enabled. Each row shows the installed version and whether that CLI is signed in. Sign in opens Terminal on that agent’s own login command; FlowDeck never handles your credentials. Enable the agents you want available, and they show up in the agent and model pickers, in Command K, and in role settings. Model List controls which of each agent’s models appear in those pickers.
Grok’s permission options are Ask and Skip Permissions only — it has no accept-edits mode to offer.

​
How it works

The agent runs through your installed CLI. FlowDeck verifies the command and its version on launch, and refreshes availability when the app becomes active after you install or update a CLI. You can see every supported provider under Settings → AI → Agents. If an installed CLI is behind the version this FlowDeck build was validated against, or newer in the catalog, an Update button appears on that row — and the same control shows during onboarding and above the chat composer when the install is below the pinned floor. Choosing Update runs the provider’s own updater in the background; FlowDeck does not open Terminal. The button reads Updating… while the process runs, then Updated or Update failed. A successful update refreshes availability immediately so the button can disappear without waiting for a refresh. Choose Refresh list, above the provider list, after installing, moving, or updating a CLI yourself to refresh immediately without relaunching FlowDeck. A temporarily missing CLI becomes unavailable but keeps your enablement preference, so reinstalling it restores your previous choice. Authentication is checked by the provider when you start using it, not during availability detection. When you send a prompt, the agent:
  1. Plans the steps needed
  2. Runs FlowDeck commands (build, test, UI automation, logs, etc.)
  3. Reads the results and adapts — fixing errors, retrying, or asking you for clarification
  4. Reports back with a plain-language summary of what it did and what it found
The agent shows its work: you can see each action it takes in real time, and the Inspector opens automatically to show issues or test results as they arrive.

​
Permissions

For actions that can’t be undone — pushing to a remote, modifying project settings, running macOS automation that affects your desktop — the agent asks for your approval before proceeding. Configure permission levels in Settings → AI (⌘,):
  • Ask for everything — approve every tool call
  • Auto-approve safe actions — builds, tests, and simulator operations run without prompting
  • Auto-approve all — fully autonomous mode; only use when you trust the full context

​
Plans

For multi-step tasks the agent may propose a plan before starting. A plan card appears in the chat showing the steps it intends to take — review it, edit if needed, then approve to start execution. The plan stays available while implementation runs, so you can keep it open for reference. When executing, Clean chat is selected by default. You can instead keep the existing chat and its context, or create a new session with a fresh chat. Creating a new session uses the usual name and project-configuration setup, then starts the approved plan automatically. The planning workspace remains unchanged.

​
Tips

Give the agent a goal, not a procedure. “Make the login tests pass” works better than “run the tests, then look at the failures, then edit the file.” The agent figures out the steps.
If the agent gets stuck or goes in the wrong direction, just say so. “Stop, the issue is in the network layer, not the UI” is enough to redirect it.
The Fix All button in the Inspector’s Issues and Tests tabs sends all current errors or failures to the agent in one step — no need to describe them.