● v.latest · macOS 12.3+ GitHub →
Context OS · ambient intelligence for builders

Sees what you see.
Hears what you hear.
Distills it into a graph you own.

A private Context OS that runs ambient in the background, distilling your screen and audio into a structured knowledge graph of facts, entities, and decisions — accessible from MCP, a web UI, and a screen-recording-invisible HUD overlay, plus peer-to-peer sharing between users.

▸ INSTALL
npx @geravant/sinain@latest start
Get started → Read the docs
licenseMIT CIpassing npm@geravant/sinain LongMemEval82.8% IPR
Demo GIF
media/sinain-demo-full.gif
loads from raw.githubusercontent.com in production
Real capture · HUD overlay over a live coding session
§ 01 — You, Augmented

Capture. Distill. Access. Share.

One knowledge graph, four uses. Sinain extracts atomic facts, entities, and decisions with an LLM — then integrates them through deterministic graph operations.

01 / capture

Screen + system audio, continuously.

<private> tag stripping and auto-redaction of credentials and tokens happen before anything leaves your machine.

macOS nativescreen + audioOCR included
02 / distill

Facts, entities, decisions.

Atomic claims, people / projects / topics, what was chosen and why. An LLM extracts — a non-LLM pipeline integrates. Moving the integration step off the LLM was the single biggest gain: 37.9% → 62.7% recall on internal Meeting Memory, 82.8% IPR on LongMemEval (ICLR 2025 public benchmark).

structured graphhybrid retrieval82.8% IPR
03 / access

MCP, web UI, HUD overlay.

Agents read it via MCP. You browse it at localhost:9500/knowledge/ui/. You glance at it through a HUD that is invisible to screen capture.

15 MCP toolslive updatesinvisible to capture
04 / share

Peer-to-peer. No server in the data path.

Send a slice of your graph to another person. Their machine pulls from yours directly over WebRTC; the data itself never touches a server. Their AI agents inherit the context via the same MCP.

WebRTCdirect transferdata stays local
§ 02 — Agent-Agnostic

One context. Any agent. No restart.

Sinain feeds the same screen and audio context to any MCP-compatible agent. Switch on the fly — no restart, no context loss.

tested with any MCP-compatible agent
Claude Code OpenClaude Codex Goose Junie
§ 03 — Privacy Controls

Four modes. From cloud-fast to fully offline.

By default, sinain uses cloud APIs (OpenRouter) for transcription and analysis. Need tighter control? Flip the mode.

▸ HUD invisible to screen capture
NSWindow.sharingType = .none
overlay/macos/Runner/AppDelegate.swift:70

The HUD never appears in screenshots, recordings, or screen shares.

Invisible to screen capture.

The HUD overlay never appears in screenshots, recordings, or screen shares.

  • off — all data flows freely; maximum insight quality.
  • standard — auto-redacts credentials before cloud APIs (wizard default).
  • strict — only summaries leave your machine; no raw text sent to cloud.
  • paranoid — Ollama + whisper.cpp. Zero network calls at runtime.
off

All data flows freely — maximum insight quality.

standard · wizard default

Auto-redacts credentials before cloud APIs.

strict

Only summaries leave your machine — no raw text sent to cloud.

paranoid · fully local

Ollama + whisper.cpp. Zero network calls at runtime.

§ 04 — Quick Start

One command. A wizard. Three services up.

On first run, sinain runs an interactive setup wizard, auto-downloads the assets it needs (~112 MB total), and starts every service for you.

~/projects · zsh
$ npx @geravant/sinain@latest start
# Pin @latest — npx caches forever against the unversioned spec.
›
▸ setup wizard — transcription · api key · agent · privacy
▸ download overlay app ~17 MB
▸ download sck-capture ~5 MB
▸ download embedding model ~90 MB
▸ install python deps
›
✓ sinain-core :9500
✓ sense_client screen + ocr
✓ overlay private · invisible
✓ sinain-agent claude-code
›
$ _

  1. Run the wizard

    Pick transcription backend, paste an API key (or skip for fully local), choose your agent, set privacy mode.

  2. Auto-download assets

    Overlay app (~17 MB), sck-capture binary (~5 MB), embedding model (~90 MB), Python deps — all cached locally.

  3. Services start

    sinain-core, sense_client, overlay, agent — running. HUD appears, screen and audio capture begins.

  4. Re-run anytime

    npx @geravant/sinain@latest start --setup. Stop / status / no-sense / headless flags available.

macOS
12.3+
Apple Silicon & Intel.
Node
18+
LTS recommended.
Python
3.10+
brew install python3, or python.org.
OpenRouter
optional
Not needed for paranoid / fully-local mode.
!
macOS Permissions. Add Terminal to System Settings → Privacy & Security → Screen Recording, then quit and reopen Terminal — macOS only applies the permission to apps started after granting. Same for Microphone.
§ 05 — Connect Your Coding Agent

15 sinain_* tools, one MCP install.

The wizard detects which MCP agents you have installed and registers sinain for each one you select.

▸ INSTALL FOR AGENTS
npx @geravant/sinain@latest mcp install

Re-run any time to update. Registers sinain with Claude Code, Claude Desktop, Cursor, Codex, Goose, and Junie.

§ 06 — Deep Dives

Threat models, audits, full configuration.

Everything that didn't fit on this page. Each link goes to the canonical doc in the repo.

Knowledge Systemdocs/knowledge-system.md → Knowledge API (HTTP)docs/KNOWLEDGE-API.md → MCP Integration (setup)docs/MCP-INTEGRATION.md → MCP Capabilities (tools + recipes)docs/MCP-CAPABILITIES.md → LongMemEval Audit · 82.8% IPRdocs/LONGMEMEVAL-AUDIT.md → Privacy Threat Modeldocs/privacy-protection-design.md → Full Configurationdocs/CONFIGURATION.md → All Hotkeysdocs/HOTKEYS.md →
Ready?

Context that follows you across every agent, every project, every machine.

One install. macOS 12.3+. MIT-licensed. Private by default.

npx @geravant/sinain@latest start Star on GitHub →