Acti wants the AI agent layer to live in your keyboard

Acti, a Singapore-based startup, has launched an AI keyboard for iOS and Android that puts agents inside an interface people use all day: the phone keyboard.

The idea is simple. Instead of copying text into ChatGPT, opening a browser, checking a stock quote, finding a restaurant, generating a meeting link, then returning to the original app, Acti wants those actions to happen from the keyboard.

It’s a smart surface. It’s also a sensitive one.

Keyboards show up inside messaging apps, email clients, social platforms, notes, search boxes, forms, and enterprise tools. If an AI assistant can work safely there, it can see intent at the moment it appears, which standalone chatbots often miss.

Acti launched on June 30 with apps for both iOS and Android. The company also disclosed a $5.3 million seed round led by BITKRAFT Ventures.

Why the keyboard makes sense for agents

Most AI assistants still force users to change context. Open an app. Paste text. Explain the situation. Ask for output. Copy it back.

That’s tedious, and it throws away useful context.

Acti’s bet is that the keyboard can act as a thin agent layer across apps. If someone asks, “Where should we eat nearby?” the keyboard can suggest a restaurant inside the chat. If a friend mentions a stock, Acti can insert a live price. If you’re writing to a colleague, it can generate a meeting link or translate a message without leaving the current app.

Founder and CEO Young Wang told TechCrunch that today’s agents are limited because user context is fragmented across separate apps. Acti’s keyboard, he argues, can create a user-owned context layer that spans them.

That’s the strongest part of the pitch. AI agents need context, but mobile operating systems don’t give third-party assistants open, persistent access to everything a user does. A keyboard is one of the few approved extension points that appears in lots of places.

It’s also one of the most privacy-sensitive.

A keyboard sees what users type. That can include passwords if apps don’t properly mark secure input fields, private messages, business discussions, medical details, financial information, and half-written thoughts that never get sent. Any AI product sitting there has to clear a higher trust bar than a chatbot tab.

Local-first is a good start, if the boundaries are clear

Acti says it uses a local-first model. According to the company, personal context stays on the device by default, and the app doesn’t access or store private messages, conversations, or personal context unless the user explicitly invokes a feature that requires external processing.

That’s the right architecture to claim for this category. Sending raw keyboard context to remote inference APIs by default would be a non-starter for many users and most regulated companies.

Still, “local-first” can cover a lot of different designs:

• Some data may stay entirely on-device.
• Some prompts may be constructed locally, then sent to external models.
• Some features may call third-party APIs for search, translation, finance, maps, or scheduling.
• Some agent actions may require authentication tokens or app permissions.

The engineering question is where Acti draws the line. Does it send only selected text? The surrounding message? App metadata? Location? User-created Skill instructions? How are logs handled? Are prompts retained by model providers? Can users inspect or delete local memory?

For consumers, Acti may be able to explain this through clear permission prompts. Developers and technical buyers will need sharper answers: data flow diagrams, retention policies, model routing, encryption details, and controls for disabling external calls.

Mobile keyboards also run under platform restrictions. iOS, in particular, limits third-party keyboards unless users grant “Full Access,” a permission that has long made privacy-conscious users wary. Android is more flexible, but that flexibility brings its own keyboard security concerns. Acti’s privacy story has to work on both platforms, not just in the product pitch.

Skills are the developer hook

Acti ships with built-in “Skills,” small user-invoked actions mapped to keyboard gestures or prompts. One example is “T,” which translates a message when the user long-presses the letter. Another is “C,” which creates a meeting link.

Users can also create Skills in plain language. Acti says early access testers built over 1,000 Skills in under two weeks before launch.

That suggests Acti is using an LLM to turn natural-language instructions into reusable workflows. The company hasn’t published deep implementation details, but the broad pattern is familiar: define intent, map it to tools or APIs, constrain the output, then expose the action through a compact UI.

The Skill model gives Acti two possible directions:

1. A personal automation layer for everyday users.
2. A marketplace for reusable micro-agents.

The marketplace path is harder. Acti plans a Skill Hub where users can share public Skills, including examples for World Cup data and Polymarket links. Marketplaces need ranking, moderation, abuse prevention, permission review, versioning, and protections against prompt-injection problems.

A public Skill that fetches sports scores is low risk. A Skill that reads a message, extracts a payment amount, drafts a reply, and inserts a link is a much bigger security and trust problem.

If Acti wants developers to take Skills seriously, it will eventually need more than natural-language creation. Plain English works for casual automation, but technical users will want inspectable logic, scoped permissions, test runs, audit trails, and deterministic failure modes. “The model decided” won’t fly when an action touches calendars, money, customer data, or workplace messages.

Distribution helps, but keyboards are unforgiving

Wang has relevant experience. He previously spent a decade at Baidu and helped grow Facemoji Keyboard to over 300 million daily active users. Acti’s CTO, Mike Sun, was the founding technical lead behind Yike Album, Baidu’s cloud-photo platform, which reportedly reached over 10 million daily active users. CSO Junbo Yang joined from HashKey Capital, where he led consumer investments.

That background matters. Consumer keyboards are brutal products. They need low latency, accurate input, good autocorrect, multilingual support, stable integrations, and a UI that doesn’t annoy people after the fifth use. AI features can’t make typing worse.

Performance will be one of Acti’s quiet make-or-break issues. A keyboard is expected to respond instantly. Even a few hundred milliseconds of lag can feel broken because users type in rhythm. If every useful action waits on a remote model call, the experience will feel like a chatbot squeezed into the wrong place.

The likely answer is a tiered system:

• Fast local behavior for typing, suggestions, and simple transformations.
• Lightweight on-device models where possible.
• Remote calls for heavier reasoning, search, translation, market data, or third-party services.
• Explicit invocation for actions that touch external systems.

That can work, but the product has to make the boundary obvious. Users need to know when the keyboard is passively assisting, when it’s reading selected text, and when it’s sending data out.

Subscriptions need obvious value

Acti plans to make money through subscriptions that offer access to more advanced AI models, higher daily limits, and premium features.

That’s predictable, and probably necessary. Inference costs money, especially if the app uses stronger models for planning or generation. Real-time data integrations and agent tooling also have ongoing costs.

The harder part is consumer attention. AI is already bundled into phones, browsers, productivity suites, and messaging apps. Apple Intelligence, Google Gemini, Microsoft Copilot, ChatGPT, Perplexity, and plenty of app-specific assistants are all competing for the same habit.

Acti’s advantage is placement. Its weakness is control. It doesn’t own the operating system or the major apps. Apple and Google can build deeper system-level AI into text fields, share sheets, notification replies, and app intents. Acti has to move faster and be more useful in the gaps those platforms leave open.

For technical teams, the subscription model also raises procurement questions. If employees install an AI keyboard that can process work messages, companies will want policy controls. Consumer virality and enterprise acceptability often pull in opposite directions.

What developers should watch

Acti is worth tracking because it points to a practical agent interface pattern: small, contextual actions embedded where users already express intent.

For developers building AI products, the lesson isn’t “build a keyboard.” Agents get more useful when they sit close to the workflow and require less context reconstruction. The best agent UI may be a text field extension, browser sidebar, IDE command palette, chat composer, spreadsheet cell, ticket comment box, or CRM note field.

The hard parts are similar across all of them:

• Capturing enough context without over-collecting data.
• Making tool calls transparent.
• Keeping latency low.
• Designing permissions users can understand.
• Handling failure safely.
• Preventing prompt injection from turning shared workflows into attack surfaces.

Acti’s Skill marketplace will be especially interesting if it grows. User-created agents sound useful until they become a distribution channel for spam, bad advice, broken automations, affiliate links, or data leakage. A good marketplace will need technical guardrails, not just community reporting.

There’s also a product-design constraint. The keyboard has limited space. If Acti fills it with buttons, prompts, icons, and agent suggestions, it risks making the most basic phone interaction feel cluttered. The best version will probably stay quiet most of the time and appear only when the user asks.

That restraint is difficult. AI products like to interrupt.

A promising surface with a high trust bar

Acti’s idea is compelling because it meets users at the point of intent. That’s where many AI assistants are weakest today. A keyboard-level agent can reduce app switching and make simple tasks feel immediate.

The same placement creates serious privacy, security, and UX pressure. A keyboard with agents can be useful. A keyboard that feels nosy, slow, or unpredictable will get deleted fast.

The company has funding, relevant consumer-product experience, and a concrete release on both major mobile platforms. Now it has to prove that agentic input can be fast, private, and boringly reliable.

For a keyboard, boring reliability keeps everything else alive.