Acti's Agentic Keyboard Explained: How Gemini-Powered Skills Work Across Apps, and the Privacy Questions to Ask First
Singapore startup Acti launched an AI keyboard for iOS and Android that runs Gemini-powered Skills, one-key shortcuts for multistep tasks, inside any app. What it does, why the keyboard is a strategic place for AI, how iOS and Android sandbox keyboards, and the data, prompt-injection and action-scope risks to check.
Acti, a Singapore-based startup, has launched an AI keyboard for iOS and Android that brings agent features into whatever app you are typing in. Powered by Google’s Gemini models, it lets users trigger “Skills”, one-key shortcuts for multistep tasks such as translating a message or sharing a meeting link, without switching apps (TechCrunch). The keyboard is a clever place for AI because it is present in every app, but it is also the most privacy-sensitive software on a phone. Acti says it is built around a local-first model that keeps personal context on the device by default. Users should still check data handling, on-device options and action confirmations before granting full access.
Acti at a glance
| Item | Detail |
|---|---|
| Company | Acti, based in Singapore |
| Founder and CEO | Young Wang, previously at Baidu, where he grew the Facemoji Keyboard |
| Platforms | iOS and Android |
| AI model | Google Gemini |
| Key feature | Skills: programmable keys that trigger multistep tasks; private or shared via a marketplace |
| Privacy approach | Local-first: personal context stays on the device by default, according to the company |
| Business model | Planned subscriptions for more advanced models, higher daily limits and premium features |
| Funding | $5.3 million seed round led by BITKRAFT Ventures |
Source: TechCrunch; company site openacti.com.
How it works
Acti replaces the phone’s default keyboard. Wang told TechCrunch that today’s AI agents are limited because user context is fragmented across separate apps, forcing people to switch apps to get AI help. A keyboard sits beneath every app, so it can help wherever you type: drafting a reply in a different tone, translating a message, sharing a meeting link in a chat, or dropping a live stock price into a conversation where a company is mentioned.
Wang said Gemini was chosen for its balance of intelligence, speed, reliability, multilingual performance and cost efficiency, all of which matter when AI is invoked during typing.
Skills: macros powered by a language model
Skills are the core feature. Users program a single key to run a multistep task automatically. The app ships with built-in Skills, such as a translation shortcut triggered by long-pressing the letter “T”. According to TechCrunch, early-access testers built more than 1,000 Skills in under two weeks. Skills can be kept private or shared to a marketplace, where examples include real-time sports data and prediction-market links.
In effect, Skills are keyboard macros with an AI model in the loop: flexible enough to handle varied input, but also capable of mistakes a rigid macro would not make.
Why the keyboard is strategic
| Layer | Examples | Advantage | Limitation |
|---|---|---|---|
| Operating system | Apple Intelligence, Gemini on Android | Deep system access | Controlled by platform owners; varies by device |
| Individual apps | Built-in assistants in email, docs and chat apps | Rich context within the app | Fragmented; each app works differently |
| Separate chatbot apps | ChatGPT, Gemini, Claude apps | Powerful and general | Requires switching apps and copying results |
| Keyboard | Acti, AI features in Gboard and SwiftKey | Present in every app without integrations | Sees everything typed; sandboxed by the OS |
Keyboard replacements such as SwiftKey and Gboard succeeded by improving prediction and typing speed. Acti’s aim is different: to remove the need to leave the app at all. The keyboard also avoids a weakness of voice assistants, which depend on app-by-app integrations; it works on the text already in front of the user.
How keyboard apps are sandboxed
Third-party keyboards run under tight operating-system limits, which shape what an AI keyboard can do:
- iOS: Custom keyboards cannot reach the network unless the user turns on “Allow Full Access”. Apple’s developer guidance says keyboards must work without it and that open access carries privacy responsibilities (Apple Developer: Custom Keyboards).
- Android: Input methods are declared as input method services, and the system warns users when they enable a new keyboard that it can collect everything they type (Android Developers: Input method).
In practice, any keyboard feature that calls a cloud model needs network access, so on iOS it will only work once the user grants Full Access. That is a meaningful decision: it lets the keyboard developer’s code send data off the device. Read the vendor’s explanation of what Full Access is used for before turning it on, and turn it off again if you stop using the AI features.
Risks specific to agent keyboards
A keyboard sees passwords, messages and payment details. Adding an AI agent raises three concerns:
- Data flow: which keystrokes or screen content go to a cloud model, and how long they are kept.
- Prompt injection: text the agent reads, such as a message from someone else, could contain instructions that manipulate it. OWASP ranks prompt injection as the top LLM application risk (OWASP).
- Action scope: whether the agent can send messages or make purchases without explicit confirmation.
Questions to ask before installing an AI keyboard
- What leaves the device? Which keystrokes, messages or screen content are sent to a cloud model, when, and how long they are retained.
- Is there an on-device or local mode, and what features work without network access?
- Does it disable itself in secure fields such as passwords and payment details?
- Does every action require confirmation before sending a message, sharing data or making a purchase?
- How are shared Skills vetted? Marketplace Skills written by others could behave unexpectedly or be designed to extract data.
- What does the privacy policy say about training on user data and sharing with model providers?
Our analysis
Putting agents in the keyboard is a smart distribution move, because the keyboard is the one interface present in every app, and Acti’s local-first design addresses the obvious privacy objection on paper. The trust bar is still higher than for any chatbot: a keyboard sees passwords, private messages and payment details. Shared Skills add another layer of risk, and prompt injection through incoming messages is a real concern for any agent that reads untrusted text. For more on that class of attack, see our guide to AI browsers and prompt injection. Users should check the vendor’s data practices, on-device options and action confirmations before granting full access, and businesses should decide whether AI keyboards are allowed on managed devices.