Claude Code is extraordinaryThe 1980s era terminal it lives in is not
Sparkle wraps the real Claude Code terminal in a modern voice-first interface built for people who think in ideas, not commands.
Just say Hey Sparkle and I'll start listening as you talk.
SendShip modern software without being stuck in the 1980s.
The 80s called. They want their terminal back.
The terminal hasn't changed since 1987. Same black box. Same cursor. Same complete indifference to how humans actually think and work.
You can't talk to it. You have to memorize arcane commands. You can't click to place your cursor. You can't highlight a word and delete it. There's no undo. There's no right-click. If you press the wrong key mid-prompt, the whole thing is gone.
Really? This is supposed to be how we interact with the most modern, powerful AI technology humanity has ever built?
The AI inside that terminal is extraordinary. The interface around it is a 40-year-old relic. Sparkle fixes that with a beautiful modern voice-first Composer interface that communicates with the 80s so you don't have to.
1980s terminal
Sparkle Composer
Software starts with an idea. Not a prompt.
Every other tool drops you straight into building. Type a prompt, get some code, hope for the best. Sparkle starts somewhere more honest: a Thinking space where ideas breathe.
In Thinking mode, triggered by the button, nothing you write touches your codebase. You're brainstorming, refining, turning a vague goal into a real spec. When you're ready, hand it to your Build agents with one click. And when your agents hit something ambiguous, you pull that question back into Thinking mode, work through it, and send the answer back. Think → Build → Think → Build. That's how real software gets made.
Thinking Mode
A safe room for ideas. Nothing you write here touches your codebase until you're ready. Chat, brainstorm, refine, spec. When you have something worth building, send it.
Build Mode
Real Claude Code agents, running in real git worktrees, building real software. Each agent names itself after what it's working on. You watch. You approve. You ship.
Thinking
Redesign auth so sessions rotate tokens on every request.
Spec: rotate on refresh, 30-minute TTL, write each attempt to the audit log.
Build
sparkle/agent-3f2a · 4 files
Just say “Hey Sparkle.”
Voice is 3x faster than typing. Stanford proved it.
Say “Hey Sparkle” and the Sparkle Composer starts listening. Speak your instruction, as long as you want, in plain English. The Composer fills with your words for review. Say “send it” and your agents have their instructions. Never auto-sends without your approval. On-device, private, $0 per word.
A peer-reviewed study from Stanford, the University of Washington, and Andrew Ng found speech runs at 153 words per minute vs. 52 for keyboard, 2.93x faster, with 20% fewer errors. When you're building, you should be talking. Read the Stanford study →
On-Device. Private.
Powered by Parakeet-TDT running locally on your Mac. Your search index and voice never leaves your machine unless you want it to.
Say "Hey Sparkle"
Always listening in the background. Mute it with one click on the waveform. The mic icon in the sidebar shows you the current state at all times.
Review Before Send
Dictation fills the Composer for review. You see exactly what will be sent before it goes anywhere. Sparkle never sends without your explicit “send it.”
Your prototype deserves to grow up.
Stop playing with toys in the sandbox and ship real production code.
Bolt, Lovable, Replit, v0 are impressive. All browser-based sandboxes. All running their own AI that is not Claude Code. When your prototype hits authentication, a real database, or production, it falls apart. Every one of them gets you to the wall and leaves you there.
Sparkle doesn't sandbox you. Your code lives in your actual file system, in real git, with real commits. You're using the same Claude Code that professional developers use. You're just not using it from a blinking cursor in a black window.
| Feature | Sparkle | Bolt | Lovable | Replit |
|---|---|---|---|---|
| Runs real Claude Code CLI | ||||
| Native desktop app (not a browser tab) | ||||
| Your real local files (not a cloud sandbox) | ||||
| Multiple agents running in parallel | ||||
| Git worktrees (agents can't step on each other) | ||||
| Voice-first input | ||||
| Vision input (drop in a screenshot) | ||||
| Mobile companion with push approvals | Yes, Soon | |||
| Open source client (MIT) | ||||
| No monthly subscription | Yes, $99 one-time | No, $25/mo | No, $25/mo | No, $20/mo |
| Where competitors win (honest) | ||||
| Browser-based, no install | ||||
| Windows and Linux support | No, macOS only | |||
| Built-in hosting and deployment | ||||
| Free tier | ||||
Sparkle requires macOS and a Claude Code install. That's the trade. What you get back is the real thing.
Everything stays private on your Mac. Until you decide otherwise.
Sparkle's base application is completely local. Your code, your prompts, your voice — nothing leaves your computer. The client stores your prompt history in local storage. Voice transcription runs on-device with Parakeet-TDT, a best-in-class open-source model that never touches the internet. Even your credentials live in your macOS Keychain. Zero cloud calls. Just you, your terminal, and Claude Code running on your machine.
AI enhancements are a different layer entirely — optional cloud features you activate individually, powered by credits. Turn on just the ones you want. Turn them all on. Turn them off mid-session. When you buy Sparkle for $99, you get $200 in AI enhancement credits to start. When you run out, re-up anytime. Use less, pay less. Use more, get more.
Sparkle macOS app: $99 · 100% local · no credits required
You will receive $200 in AI Enhancement credits when you buy the Sparkle macOS app.
| Feature | 1980s terminal | Sparkle macOS clientPrivate · local | Sparkle + AI enhancementsPay-as-you-go with credits |
|---|---|---|---|
| Compose & prompt | |||
| Modern terminal editor | Not available. Stuck in the 1980s | Available. Click anywhere, select, delete, rewrite, undo/redo, every standard shortcut (⌘C/V/Z/A), and minimize to the terminal with ⌘J. | |
| A picture is worth 1,000 words tokens. Give Claude Code vision. | Not available. | Available. Local only. Camera → crosshair → thumbnail in the composer. Stored locally. | |
| Prompt history, autocomplete & Time Machine | Not available. | Available. Local only. Re-use past prompts, autocomplete as you speak or type, and jump back to any moment in your terminal (last 24h, local). | Needs AI credits. Up to a month locally, with optional cloud backup to keep it indefinitely. |
| Search your terminal history | Not available. | Available. Local only. Keyword search the last 24h of sessions. | Needs AI credits. Semantic (by meaning, not just text) + keyword, up to a month locally, with optional cloud backup. |
| Voice input | |||
| Voice-first dictation, always listening | Not available. | Available. Local only. On-device Parakeet-TDT. Always listening (one click to turn off); say “Hey Sparkle” to start, “send it” to finish. Fills the composer, never auto-sends. | Cloud voice for realtime streaming, higher accuracy, noise filtering, less word confusion (“write” vs. “right”), and auto re-wording when you correct yourself. |
| Orchestrate & build | |||
| Run many real agents in parallel | Not available. One at a time, unmanageable | Available. Local only. Multiple tabs, each a real PTY in its own isolated git worktree (fresh origin/main, no collisions). | Needs AI credits. An orchestrator reads your spec and spins workers up and down automatically. |
| Keep track of every agent at a glance | Not available. | Available. Local only. Color-coded status dots + attention-aware ordering (red / green / gray) so you see who needs you, plus staleness badges before agents reason on stale code. | Needs AI credits. Auto naming & re-naming of every agent tab so you remember what each one is doing. |
| Status pipeline: uncommitted → committed → PR → merged | Not available. | Available. Local only. Live, from local git status. | Needs AI credits. Adds a production stage, gated by the Readiness Gate. |
| Understand & navigate the terminal | |||
| See everything Claude Code does, in a modern UI | Available. That's all you have | Available. The same terminal, wrapped in a modern UI. | |
| Highlight any terminal text for instant actions | Not available. | Available. Local only. Copy, run as a command, search the web, send to your agent, or save a note/task to the built-in issue tracker — all local. | Needs AI credits. Plus AI actions: Explain this, Fix it, or Think it through. |
| Stay in control (approvals & safety) | |||
| See, and approve, what your agents do | Not available. You watch the cursor | Available. Local only. A local classifier risk-tiers every action (safe / caution / dangerous); agents that need you float to the top with a red dot; approve or reject in-app. | Needs AI credits. Also pushed to your phone — approve from the iOS app; dangerous actions interrupt + push + SMS + email (coming soon). |
| Production Readiness Gate before any deploy | Not available. | Not available. | Needs AI credits. A 0–100 score across security, quality, and completeness. Below 50 blocks the deploy (configurable). |
| Safe by default | Not available. Easy to do the wrong thing | Available. Local only. No direct pushes to main: PRs enforced (configurable); your credentials stay in the macOS Keychain, never logged or sent. | Available. Local only. Sparkle never sees your API key in plaintext. |
| Think & plan | |||
| Thinking mode: brainstorm, then build | Not available. | Not available. | Needs AI credits. A safe space that never touches your code — refine an idea into a spec, then hand it to a Build agent in one click. |
| An AI that understands your whole project | Not available. | Not available. | Needs AI credits. Build agents sync their progress to the Thinking context and flag ambiguity back to you — a bidirectional Think → Build → Think loop. |
| Privacy & data | |||
| Your code and prompts stay on your Mac | Available. | Available. Local only. Source code is never uploaded; prompts live in local storage and are never transmitted. | Available. Local only. Cloud features never access your source code; search sends only embeddings. |
| Voice never leaves your machine | Not available. N/A | Available. Local only. Parakeet runs on-device. | Optional cloud voice; base stays local when off. |
| Credentials in Keychain | Not available. N/A | Available. Local only. API key in your macOS Keychain (never logged or sent). | |
| Open source — audit every line | Not available. N/A | Available. All client code is public (MIT). | |
$ claude code
● Waiting for your instruction…
Fix the padding on the login modal header
The terminal is still there. You just don't have to live in it.
The terminal hasn't changed since 1987. You can't click to place your cursor. You can't highlight a word and delete it. There's no undo. There's no right-click. Accidentally press the wrong key mid-prompt and the whole thing is gone. This is the interface that stands between you and the most powerful AI coding tool ever built.
The Sparkle Composer is what the terminal's input always should have been. A rich, modern editor where you can click anywhere, highlight anything, drag, drop, and rethink before you send. Drop in a screenshot and Claude Code can see exactly what you see. The terminal gets a perfect instruction. You got to think it through first.
Give Claude Code Vision
Hit the camera button, drag a crosshair over any part of your screen, and drop it in. Claude Code sees your UI bug, your design mockup, your error dialog. No copy-paste. No description. Just show it.
Ghost-Text Autocomplete
Every prompt you've ever written is saved. Start typing and your history completes it. Tab to accept. The longer you use Sparkle, the faster you build.
Minimize + Restore
Press ⌘J or drag the handle down to tuck the Composer into a slim 22px bar. The terminal comes full-screen. Hit ⌘J again and your Composer snaps back. You're always one keypress from your next instruction.
Run two agents. Approve from your couch.
Start an orchestrator. Give it a goal. It reads your spec and spins up workers automatically, each isolated in its own git worktree, each working on its own branch, each named after exactly what it's building. They can't step on each other. They can't collide. They can't merge into main without your say.
Agents that need a decision float to the top. Agents that are happily building sink to the bottom. You're not watching a terminal blink. You're running a software operation.
2 orchestrators · 5 workers
Approval requested
🔴 Dangerous
Drop the staging database?
DB schema migration wants to run an irreversible command before rebuilding the schema.
Push · SMS · Email sent 2s ago
Your agents work while you live your life.
You don't need to be at your desk. Sparkle knows when something needs your attention and sends it to your phone. A blocked agent waiting for your approval. A judgment call. A risky action. Sparkle sends push notifications to your phone. Tap to see what your agent wants to do. Tap Approve or Reject. Done. Your agent keeps building.
Three tiers of escalation. Caution: in-app + push. Dangerous: push + SMS + email all at once. Morning briefing: a summary of what your agents built overnight, to read while you drink your morning coffee.
Caution
Agent flagged this as risky
Push + in-app
Dangerous
Wants to do something irreversible
Push + SMS + Email
Build Complete
Your session finished
Push + Email
Morning Briefing
Overnight build summary
Email + Push at 7am
We'll notify you when the iOS app is ready. No spam.
Every guardrail is something you'd otherwise forget.
Sparkle is built around a set of strong opinions about how software should be built with AI agents. These opinions are what make it work for people who just want to build without reading a 40-page guide on git workflows. Here's every opinion Sparkle enforces, and why.
| The Rule | What It Prevents |
|---|---|
| No direct pushes to main: PRs required | Agents accidentally shipping broken code to production |
| Every agent starts from a fresh origin/main | Building on top of another agent's unfinished work |
| No force-pushing to main, ever | Silently dropping commits that other agents already built on |
| Production only deploys through main | Accidental shipping of untested feature branches |
| Risk-tiered approvals: caution → dangerous → auto | Agents deleting your database without your knowledge |
| Production Readiness Gate (0–100 score) | Shipping something half-built because an agent said it was done |
| Agent worktrees isolated outside the project tree | An agent writing to files it wasn't supposed to touch |
| Your credentials never pass through Sparkle | Your API keys being captured in logs or training data |
The full list of 16 enforced rules, and which ones open source users can modify, is on the Features page →
Highlight anything. Understand everything.
The terminal is visible. You can see everything Claude Code is doing. But you don't have to decode it alone.
Highlight any text in the terminal (an error, a command, a file path, a cryptic stack trace) and get a menu of ten actions: Explain this. Copy it. Search the web about it. Run it as a command. Create a task from it. Save it as a note. You're not expected to know what any of it means. That's Sparkle's job.
You've been a builder this whole time.
You just didn't have the modern tool you needed to build your thoughts.
macOS 11.0+ · Universal · ~7MB · No monthly subscription
iOS app coming soon. Join the waitlist