Simply speak to launch any number of specialized AI agents to perform tasks in real-time on any screen, application, or website on your Mac.
Instantly spawn an agent to:
Ditto sees and adapts to anything on your screen in real-time, guiding or taking control depending what you request.
Hands-free, voice-first to perform anything instantly on your desktop.
Voice command to first cursor move on your screen.
MCP tools connected — Mail, Calendar, browser, IDE, every app you use.
Works with Claude and OpenAI plans
YouTube Tutorial Agent
Watches → mimics on your Mac
Browser Agent
Find, fill, click — voice-driven
Workflow Agent
A starter agent that captures and replays workflows across your installed apps.
Workflow Agent
Build once, replay anywhere
Voice
On-device wake-word detection — no audio sent to the cloud until you confirm. Ditto transcribes locally, snapshots your live screen, and reasons about the request before any agent moves.
Wake-word → STT → agent reasoning → first cursor moves. The full voice loop in 30 seconds.
demo-wake-word-flow.mp4
"Hey Ditto" detected by an on-device model. Audio never leaves the Mac until the wake word fires.
Ditto snapshots what's visible (with privacy redaction) and the transcript becomes the agent's task brief.
A planning agent picks which scoped sub-agents (Mail / Calendar / Browser / Research / Notes / Summary) handle the work. You see them go.
When the agents finish, Ditto synthesizes the result and reads it back. ElevenLabs TTS or system voice — your pick.
Shadow Cursors
Every Ditto sub-agent gets its own colored shadow cursor on screen. Not abstractions — actual pointers you can watch click, type, and read across native apps and web tabs. The cursors moving on this page are doing exactly what they'd do on your Mac.
Speak a goal — Ditto chains agents across Mail, Calendar, Slack, your IDE. Captures the workflow so it replays on demand.
Live shadow cursors point, click, and explain as you onboard. Ditto reads the screen and shows you what to do.
Spawn parallel sub-agents — one drafting in Notion, one cross-referencing your inbox, one updating your calendar. All at once.
Validate user flows across your real app, browser, and APIs — Ditto runs the test like a person would.
Reads, drafts, sends, archives in Apple Mail or Gmail.
Schedules, reschedules, finds free slots.
Multiple shadow cursors moving in parallel across the desktop — one per sub-agent — color-coded by role.
demo-shadow-cursors.mp4 · screenshot-shadow-cursors.png
Ditto Skills
Ditto is more than a thin wrapper around MCP servers. Speak intent and Ditto builds workflows, mentors you through unfamiliar apps, runs end-to-end UX tests, and streams structured telemetry — all from your voice.
Speak a multi-step task and Ditto builds a live workflow. Linear → Slack → Calendar in one prompt. Saved for next time.
demo-workflow-builder.mp4
Stuck in Concur, Workday, anything? Ditto takes the cursor, walks each click with annotations, hands control back.
demo-show-me-how.mp4
Hand Ditto a flow. It clicks, types, navigates, verifies — like a real user. QA, regressions, demo recordings, onboarding paths.
demo-e2e-testing.mp4
Every shadow-cursor action streams structured telemetry to Loki / DataDog / Honeycomb / your webhook.
demo-telemetry-stream.mp4
MCP-Connected Tools
Ditto speaks Model Context Protocol natively. Any MCP server you connect — Mail, Calendar, Browser, Files, Contacts, Slack, Linear, GitHub — becomes a tool every sub-agent can call. No bespoke integrations to maintain.
Apple Mail + Gmail. Sends always behind explicit approval.
Google + iCloud. Find slots, schedule, decline.
Chrome MCP. Forms, scrapes, multi-step flows.
Finder + Spotlight. Open, move, rename, search.
Reads any open app on your Mac via Accessibility API and Screen Recording — clicks, types, navigates like a person.
Slack, Linear, GitHub, Notion, Stripe — drop in.
"Hey Ditto, draft a reply to my last email from Sarah, and find a 30-min slot tomorrow." Watch the Mail and Calendar cursors split the work in real time.
demo-mail-calendar-mcp.mp4
Product Surfaces
Ditto's notes & memory surface — captured snippets, retrievable by voice, filed by tag.
demo-notes-memory.mp4 · screenshot-notes-memory.png
Ditto performing native macOS actions — clicking, typing, scrolling — driven by voice intent.
demo-desktop-execution.mp4 · screenshot-desktop-execution.png
Browser cursor scraping product specs across multiple tabs, condensing into a doc.
demo-browser-research.mp4
License + Pricing
Ditto is sold as a desktop license. You pay a flat monthly rate for the app + sub-agent runtime; you bring your own Anthropic / OpenAI / ElevenLabs keys for model calls.
For one operator on one Mac.
Workspaces, approval policy, audit trail.
Pricing is a placeholder; subject to change before public launch. Contact for enterprise.
FAQ
Yes. Ditto is bring-your-own-keys for model calls (Anthropic Claude, optionally OpenAI, ElevenLabs). Paste them in Settings on first run; they're stored in macOS Keychain and never leave your device except as the request body to those providers.
macOS Accessibility (so shadow cursors can move + click), Microphone (wake word + push-to-talk), Screen Recording (live screen context), Speech Recognition (on-device wake word), Apple Events (open Settings deep-links during onboarding). All standard TCC prompts; revocable any time in System Settings → Privacy & Security.
Ditto detects the revocation and immediately suspends wake-word + push-to-talk capture. The menu-bar status switches to "Mic disabled" and a banner tells you how to re-grant. Sub-agents already in flight finish; no new voice input is captured.
Today: Anthropic Claude (default), OpenAI GPT-4o (optional), with Ollama local-LLM as a future fallback for the orchestrator path. The Ditto voice loop is currently Claude-only because of vision + tool-use stability. Local-LLM voice path is on the v2 roadmap.
For now, yes — macOS 14.2+. Ditto is a native Swift app (~17K LOC) and uses ScreenCaptureKit, AVAudioEngine, on-device Speech, and the macOS Accessibility APIs that aren't portable to other OSes. Windows / Linux versions aren't planned.
No raw audio. Wake-word detection runs on-device. Once the wake word fires, your post-wake-word transcript text (not audio) goes to Claude's API along with a redacted screen snapshot. ElevenLabs receives only the response text Ditto speaks back. Set TTS to system-voice in Settings if you'd rather keep voice synthesis local too.