convez.ai
MCP ServerIntegração & APIsAvançado
Monkopedia

Rouse Context

Monkopedia/rouse-context

Acesso seguro aos dados do seu Android diretamente no Claude, sem cloud.

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP para Android que expõe dados do dispositivo (saúde, notificações, uso de apps) diretamente ao Claude sem sincronizar para a nuvem. Usa FCM para acordar o telefone sob demanda, estabelece sessão TLS criptografada ponta-a-ponta e volta ao estado de repouso quando desconectado.

Instalação

$npx convez-ai Monkopedia/rouse-context

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
14
Forks
1
Acompanhando
1
Saúde
71%
Último commit
há 1 semana
Versão
v1.0.14
Issues abertas
73
Contribuidores
2

Casos de uso

  • Consultar dados de saúde (atividade, sono, vitais) em contexto de IA
  • Automatizar envio de notificações e consultas de uso de aplicativos
  • Integrar dados pessoais do Android em fluxos de IA sem exposição de privacidade

Veredito

Solução engenhosa de privacidade para Android que resolve um problema real: dar ao Claude acesso a dados pessoais sem comprometer segurança. Arquitetura sofisticada (relay SNI, mTLS, FCM) exige conhecimento técnico, mas a implementação está madura (v1.0). Ideal para construidores que priorizam zero-knowledge architecture.

README

A tradução para o português deste README ainda está na fila.

Rouse Context

rouse-context MCP server

Your phone has context that AI doesn't — your health data, your notifications, your app usage. Rouse Context makes that context available to AI assistants on demand, without ever syncing your data to the cloud.

It's an Android app that turns your phone into an MCP server. AI clients like Claude connect to a URL, your phone wakes up, and a direct encrypted session is established. The AI asks for what it needs, your phone responds, and then it goes back to sleep. Your data never leaves your device except through that live session.

Status: v1.0 released. Available on GitHub Releases. For Google Play beta access, email beta@rousecontext.com. Works with any MCP-compatible client.

Getting Started

  1. Install — download the APK from GitHub Releases and sideload it (you may need to enable "Install from unknown sources" in your Android settings). Or email beta@rousecontext.com for Google Play beta access.
  2. Onboard — open the app, tap Get Started, grant notification permissions. The app registers with the relay and provisions your device's TLS certificates (~10 seconds).
  3. Enable an integration — tap Add Integration on the home screen. Pick one (e.g. Health Connect), grant the requested permissions, and you'll get a URL like https://brave-health.abc123.rousecontext.com/mcp.
  4. Connect your AI client — paste that URL into Claude (Settings → Connectors → Add custom connector) or any MCP client. The first connection triggers an OAuth approval on your phone — tap Approve.
  5. Use it — ask the AI about your health data, send yourself a notification, check your screen time. Every tool call is logged in the app's Audit tab.

How It Works

AI Client ──TLS──> Relay (SNI passthrough) ──mux WebSocket (mTLS)──> Your Phone
  1. You enable an integration (e.g. Health Connect) and get a URL like https://brave-health.abc123.rousecontext.com/mcp (the integration is identified by the hostname prefix; the path is always /mcp)
  2. Add that URL to Claude or any MCP client
  3. When the client connects, the relay wakes your phone via FCM push
  4. Your phone connects back through a mTLS WebSocket, and the relay splices the two TLS streams together
  5. The AI client talks directly to your phone over end-to-end encrypted TLS — the relay never sees the plaintext

Sessions are ephemeral. The phone goes back to sleep when the client disconnects.

Integrations

Each integration is independently enabled and gets its own MCP endpoint with OAuth authorization (PKCE). 22 tools across 4 integrations:

Health Connect (3 tools)

Query Android Health Connect data across 36 record types in 7 categories (activity, body, sleep, vitals, nutrition, reproductive, mindfulness).

ToolDescription
list_record_typesList Health Connect record types with permission status
query_health_dataQuery records by type and time range
get_health_summaryHealth summary across permitted types for a period

Notifications (5 tools)

Read, search, and interact with device notifications.

ToolDescription
list_active_notificationsList posted notifications
search_notification_historySearch notification history by text, package, or time range
get_notification_statsNotification counts, top apps, and busiest hour for a period
perform_notification_actionInvoke an action button on an active notification
dismiss_notificationDismiss an active notification by key

Outreach (10 tools)

Let AI take actions on the device — launch apps, open links, send notifications, manage DND.

ToolDescription
launch_appLaunch an installed app by package name
open_linkOpen an http/https URL in the default app
copy_to_clipboardCopy text to the clipboard
send_notificationPost a notification with optional action buttons and tap-to-open URL
list_installed_appsList installed apps
create_notification_channelCreate a notification channel
list_notification_channelsList AI-created notification channels
delete_notification_channelDelete an AI-created notification channel
get_dnd_stateGet Do Not Disturb state
set_dnd_stateSet Do Not Disturb state

Usage Stats (4 tools)

App usage patterns and screen time analytics.

ToolDescription
get_usage_summaryScreen time totals and top apps for a period
get_app_usagePer-day usage for one app
get_usage_eventsRaw app foreground/background events over a range
compare_usageCompare screen time between two periods; biggest deltas first

Architecture

Android App

ModulePurpose
:appCompose UI, Koin DI, navigation, theming
:core:tunnelMux protocol, WebSocket client, TLS accept, CertificateStore
:core:mcpMCP session routing, HTTP server, OAuth (device code + auth code + PKCE), token management
:core:bridgeBridges tunnel mux streams to MCP sessions via TLS
:apiMcpIntegration interface, IntegrationStateStore
:integrationsMCP providers: Health Connect, outreach (calls/SMS/DND), usage stats, notification capture
:notificationsCross-cutting notification infrastructure, audit persistence (Room)
:workForeground service, FCM receiver, WorkManager
:core:testfixturesShared fixtures for integration-tier tests that boot the real relay binary

Relay Server (relay/)

Rust binary on a small VPS. Handles:

  • TLS passthrough — routes AI client connections to devices via SNI, never terminates inner TLS
  • Mux WebSocket — multiplexes multiple client sessions over one device connection
  • FCM wakeup — sends push notifications to wake sleeping devices
  • ACME certs — issues per-device TLS certs via Google Trust Services DNS-01 (Cloudflare API); Let's Encrypt also supported
  • Bot protection — secret URL prefix validated before waking device, plus FCM throttle and IP rate limiting

Security

  • End-to-end encrypted — TLS terminates on your phone. The relay does SNI passthrough only.
  • No cloud sync — data is read from on-device sources and served live. Nothing stored remotely.
  • Per-device ACME certs — per-device certs via Google Trust Services (DNS-01), private key in Android Keystore (hardware-backed).
  • mTLS device auth — relay authenticates the device by client certificate before allowing connections.
  • Secret URL prefix — each device URL includes a rotatable per-integration secret (brave-health.abc123.rousecontext.com, where brave-health is {adjective}-{integrationId}). Bots that discover the device subdomain can't wake it without the secret.
  • OAuth per-client — each AI client must be authorized via on-device approval before accessing tools.
  • Audit trail — every tool invocation is logged locally with arguments, response, and duration.

Building

Android

./gradlew assembleDebug

Requires Android SDK (compileSdk 36, targetSdk 35, minSdk 24). Build requires JDK 21 (JAVA_HOME=/usr/lib/jvm/java-21-openjdk).

Coverage report

./gradlew koverHtmlReport

Aggregates line + branch coverage across every module's unit tests and the :core:tunnel:integrationTest tier (real relay subprocess). HTML lands in build/reports/kover/html/index.html. CI publishes the same report as a test-coverage artifact plus a per-module summary on each PR.

Relay

cd relay
cargo build --release

End-to-end tests

The e2e tests drive a real device over adb via a (typically LAN-local) host running adb. The adb.host system property is required and defaults to empty — the :e2e:e2eTest task fails fast if unset. Device serial is optional and only needed if multiple devices are connected to that host.

./gradlew :e2e:e2eTest \
    -Dadb.host=<your-dev-host> \
    -Dadb.serial=<your-device-serial>

The :device-tests module has a separate runner that builds an APK pointed at a locally-running relay. It also requires a LAN IP reachable from the device:

./gradlew :device-tests:test -Dlan.ip=<your-lan-ip>

Without lan.ip, the device-tests skip cleanly via JUnit assumptions.

Status

v1.0 released. 22 tools across 4 integrations, cold-start wake in ~3 seconds, end-to-end encrypted. See docs/design/ for detailed design documents and privacy policy.

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.

Ver o original no GitHub

Bom para

Relacionados