tvara

trytvara.com

A new way to navigate your Mac.

A new way to navigate your Mac

PS: First stable version is WIP! <3
Your messages, photos, emails, notes, browser history, and clipboard all live on this disk. To find any one of them today, you open the right app and use its searchbox. Twelve apps, twelve searchboxes, twelve different ideas of what search should do.

tvara is one searchbox across all of them.

⌘K from anywhere. Type a few words. Press ↩.

What it searches

• Apps, files, folders
• Messages: WhatsApp, iMessage, Discord
• Mail
• Notes
• Clipboard history
• Browser history: Chrome, Arc, Brave, Edge
• Photos: on-device semantic image search (CLIP)
• System Settings panes
• Window snap actions and system actions (lock, sleep, restart)

Queries can be natural-language: "the chase email about my credit card", "the photo with the dog at the beach", "snap this window top right". The launcher tracks which results you pick and weights them up within their rank band, so things you reach for often start surfacing higher over time.

Install

A signed binary is coming to trytvara.com. For now, build from source:

./scripts/build-app.sh
open ./tvara.app

Requirements

• macOS 14 (Sonoma) or later
• Xcode Command Line Tools (xcode-select --install)
• An Apple Development cert in Keychain. The build script signs the bundle with it so macOS TCC grants (Accessibility, Contacts, Full Disk Access, Automation) persist across rebuilds. Otherwise you'd re-click those prompts every time you rebuild. List yours with security find-identity -v -p codesigning and update SIGNING_IDENTITY in scripts/build-app.sh.

First launch. macOS will prompt for permissions in a batch. Grant them once and you're done. tvara summons with ⌘K from anywhere.

Keys

Key	Action
⌘K	summon
↑ ↓	move selection
↩	open
⌘↩	act on selected (send a message, create an event)
⇥	category deck
↩ (in deck)	zoom into a category
esc	back one layer (zoom → deck → blended → clear → hide)

How it works

Per-keystroke fan-out across local sources: files via Spotlight's own mdfind, messages via direct SQLite reads from each app's local store, photos via on-device MobileCLIP semantic embeddings, mail via Apple Mail's FTS index. Results stream into a single ranked list as each source returns. You see the fastest hits in milliseconds while the slower ones land underneath.

An optional natural-language planner (OpenAI today, on-device next) routes ambiguous queries. For example, "address i sent sam last week" gets parsed into source: messages, contact: sam, time: week, search_term: "street address", so the right source gets the right query without you thinking about it.

A frequency reranker watches which results you actually pick and weights them up within their rank band. The launcher gets sharper the more you use it.

Privacy

Your data is yours. tvara is built to keep it local by default. We think that's table stakes for any tool that touches your messages, photos, and mail.

What stays on your Mac. All search across apps, files, messages (WhatsApp / iMessage / Discord), Mail, Notes, browser history, clipboard, and System Settings runs locally against each app's own database or via macOS APIs. Photo semantic search uses MobileCLIP running on-device via CoreML, so images never leave your disk. The frequency reranker stores selection counts in ~/Library/Application Support/tvara/ as local SQLite, never synced.

What touches the cloud. Two narrow paths use OpenAI today. Both are BYOK (bring your own key; we don't proxy, broker, or aggregate keys) and only fire when invoked:

• Natural-language query planner (gpt-5.5). When you type a query like "the chase email about my credit card last week," tvara sends just the query string to OpenAI to parse out source + contact + keywords + time. The structured plan comes back, and tvara searches your local data. Your actual messages, files, photos, and results never leave the machine.
• Compose-action planner (gpt-5.5). When you ⌘↩ on a result to draft a message or calendar event, tvara sends your action intent plus the snippet you're acting on (capped at 800 characters). Nothing else.
• Discord semantic rerank (text-embedding-3-small). For some Discord queries, tvara embeds the planner's distilled search term (e.g. "street address") so it can rank pre-computed message vectors. Never raw message content. This is a placeholder. We're training a small on-device embedding model to replace OpenAI on this path; once it ships, even this distilled phrase stops leaving your Mac.

Without an API key configured, none of these fire. Natural-language planning silently falls back to literal keyword matching against your local data. The cloud paths are an opt-in upgrade, not a requirement.

Local models, eventually. The plan is to swap the cloud planner for an on-device LLM. Honest tradeoffs today:

• Local 3B models via Ollama work, but our benchmarks (experiments/bench_local_planner.py) put them at ~3s per query vs OpenAI's ~500ms-1.5s.

• Smaller models classify the source correctly ~80% of the time vs near-100% for OpenAI.

• Apple's FoundationModels framework (on-device, free, no network) is the cleanest path forward. We benched it against OpenAI on the compose-action planner (experiments/bench_planner.swift, 5 samples × 3 runs on an Apple Silicon Mac running macOS 26.5):

  • openai/gpt-5.5 (low reasoning): avg 3985 ms, p50 3585 ms, p95 5588 ms — 100% type-accuracy, 100% shape-accuracy.
  • apple/foundationmodels (on-device): avg 1633 ms, p50 1578 ms, p95 3055 ms — 100% type-accuracy, 93% shape-accuracy (one missing field on a long input; expected to pin at 100% once we move to the @Generable structured-output path).

  Roughly 2.4× faster than OpenAI on average, free, works offline. Requires macOS 26 + Apple Intelligence enabled, so it can't be the default yet — but it's the obvious win once we can.

• Performance varies meaningfully by system: Apple Silicon vs Intel, RAM headroom, which model you've pulled.

So today: OpenAI is the default with BYOK. Anthropic/Claude as another BYOK option is on the list. Local-LLM swap will ship as an opt-in once the latency story is good enough on enough machines.

Telemetry. None. tvara doesn't phone home, doesn't track usage, doesn't ping for updates, doesn't ship analytics. The only outbound network calls are the OpenAI ones above, and only when you've provided a key.

Status

Active development. Some sources (Notion, Linear, Spotify) require API keys; others (Messages, Mail, WhatsApp, Discord, Notes, Clipboard, Photos) work out of the box once macOS permission prompts are granted on first launch. The architecture is set; the polish is in flight.

License

PolyForm Noncommercial 1.0.0. Free for personal, hobby, research, and non-commercial-organization use. Commercial use reserved. Commercial licensing: aaryanshsahay7@gmail.com.