Intent Keyboard
January 28, 2026You have an idea. You want it done. But first: open the app, navigate the menu, configure the thing, wait for it to run.
What if you could just press a key?
Try it
Press spacebar. Watch the command execute. See the text change.
The intent keyboard transforms how we interact with software. Press one key, and an AI agent figures out what you need.
One key
Multiple keys mapping to different agents breaks down. Intent is fluid. Context is heavy. You end up maintaining a mental model of which key does what in which situation.
One key is simpler. Press it, and an AI agent gets full context: screen state, active app, recent history, project state. The agent decides what action makes sense right now.
The demo
You just committed auth changes. Press the key.
preview-auth-7f3k.vercel.appIf tests fail: creates ticket, assigns you, links to logs.
One key with verbs
The key doesn't represent an app. It represents a contract: "Take over and move me forward." Different gestures give you different verbs.
TapDo the next obvious actionDefault best guess based on context
HoldExplain / summarize / proposeNo execution, just show options
Double-tapRepeat last successful intentSame action, fresh context
Shift + tapLocal onlyNo network, no external APIs
Option + tapDry runShow plan, require confirmation
Cmd + tapComposeDraft output: email, PR, ticket, message
Agency ladder
Trust is what kills agent buttons. You need to know what it can do and when it asks permission.
Shows what it would do. No changes.
Creates output. You approve before send.
Runs actions that can be undone.
Always confirms first. Deletes, payments, external posts.
Timeboxed, cancellable. You set the window.
Receipts
Every action has a receipt. You can see what it saw, what it decided, what it changed, and how to undo it.
vercel rollback preview-auth-7f3kSame key, different context
The agent adapts. Same key, different situation, different action.
Not macros
Macros execute fixed sequences. This reasons about context and chooses the right action.
if (keypress) {
run("deploy.sh")
}context → reasoning → action Adapts to what you're actually doing
The bet
The keyboard is a wedge. The real product is the context layer (what's on screen, in repo, in tools), the policy layer (constraints, approvals, audit), and the workflow runtime (tools, connectors, retries, state).
If pressing the key reliably does something high-leverage with low latency and high trust, the concept isn't pointless. Software becomes something you command, not something you operate.