Sutrana

Local-first ambient memory for Mac: continuously captures what you see, indexes it on-device, exposes it as an MCP server for Claude/Cursor, and surfaces open loops in a Shram-style menu-bar inbox.

Status: Early public preview (v0.1). Capture → SQLite/FTS → MCP → menu-bar open tasks works on Apple Silicon with full Xcode. Not a signed App Store build yet.

What it does

• Menu bar app — open tasks stream, settings, permissions, kill switch
• Screen capture + OCR — ScreenCaptureKit + Vision, local SQLite + FTS5
• MCP tools — search_memory, get_timeline, list_open_tasks, suggest_followups, list_day_summaries, embed_recent_memory
• Optional hybrid search — local Ollama embeddings (off until you run Ollama)

All data lives under ~/Library/Application Support/Sutrana/. No cloud by default.

Quick start

1. Memory + MCP (Node)

git clone https://github.com/sdntsng/sutrana.git
cd sutrana
npm install
npm run build
npm run mcp

2. Mac menu bar app

Requires full Xcode (Command Line Tools alone are not enough).

cd apps/mac
./Scripts/build-app.sh
open Sutrana.app

Grant Screen Recording (and optionally Accessibility) when prompted — or use the Permissions tab.

3. Connect Claude Desktop / Cursor

See docs/claude-desktop.md and docs/cursor.md.

Privacy

• Local-only storage and embeddings by default
• Private-apps denylist in Settings
• Pause capture anytime from the menu bar
• Microphone / accessibility scrapers are opt-in later; mic is not used in v0

See SECURITY.md to report issues.

Repo layout

apps/mac/           Swift menu-bar capture agent
packages/memory/    SQLite FTS + open tasks + day summaries + Ollama embeddings
packages/mcp/       MCP stdio server
docs/               client config examples

Development

• Default contribution branch: dev (staging)
• main is production / releases
• npm run build · npm run typecheck · npm test

See CONTRIBUTING.md.

Roadmap (post-v0.1)

• Signed .dmg distribution
• Stronger open-task heuristics / LLM drafts
• Accessibility messaging scrapers (opt-in)
• Optional remote embeddings with explicit consent

License

MIT