Built for phones, not for demos
The details that decide whether an agent is trusted with your real apps.
Features
Everything between “I want” and “it's done”
An agent that acts, a console that keeps you in control, and a way to make good runs repeatable.
Natural-language control
Say the goal. The agent sees the screen and the UI structure, then taps, types, swipes and checks the result.
Human takeover
Watch the phone live in the browser and take over mid-run. Hand it back and the agent carries on with what you did in mind.
Flows you can replay
Turn a successful run into a readable Flow. Replays take seconds and cost nothing in model usage.
Scheduled tasks
Run a goal at set times each day, tried out first. Repeated failures pause it automatically.
Background screens
Run the agent on a virtual screen so your phone's main screen stays yours.
Bring your own model
Use a built-in provider or any OpenAI-compatible gateway. DuckClick tests whether a model can see images and call tools before you rely on it.
App guides and skills
Short per-app notes and installable skills teach the agent the traps of your apps, so it stops repeating mistakes.
Guardrails built in
App policies, approvals, action budgets, cost caps and a full event log. It never bypasses a PIN, pattern or password.
An agent that verifies
Every tap is followed by a fresh look at the screen. Typing is checked against the focused input. If an action changes nothing, the agent is stopped from repeating it forever.
- Screenshot plus UI hierarchy on every step
- Typed text is verified before pressing Enter
- Repeated no-op actions are warned, then refused
You and the agent, one phone
Only one operator touches a phone at a time, enforced by the server. A person can take over, and what they did is recorded so the agent understands the new state.
- Live screen in the browser, with touch control
- Take over and hand back without losing the conversation
- Per-phone queue when several jobs want the same device
Flows: explore once, replay forever
A Flow is a short YAML file distilled from a successful run. The interpreter replays it directly on the phone, with no model in the loop.
- Single-step Flows run in roughly 2 to 6 seconds
- If an element is missing the run fails fast and lists what is on screen
- Flows can be edited, versioned and re-verified
Unattended, but not unsupervised
Scheduled tasks run on their own, yet stop and report when a human is needed: a lock screen, a login, a captcha.
- Daily schedules, tried out before they go live
- Pause after consecutive failures
- One switch pauses everything
Any model, your choice
Pick a provider or point DuckClick at your own gateway. It probes chat, tool calling, vision and reasoning, and only trusts the models that pass.
- OpenAI-compatible gateways, one-api, vLLM, Ollama and more
- API keys are write-only in the console
- Per-run cost cap
Give your phone a duck to do the tapping.
Connect a phone, say what you want, and watch it get done.