J.A.R.V.I.S. Protocol

The Jarvis paradigm — an always-available, context-aware, voice-enabled personal AI — is no longer science fiction. Claude Code, combined with its MCP server ecosystem, headless mode, and remote control capabilities, provides every building block needed to construct a personal assistant that rivals the fictional original. This report synthesizes research across six domains: remote mobile access, voice interaction, messaging bridges, home automation, daemon orchestration, and the open-source tooling landscape.

Five distinct pathways exist for accessing Claude Code from a phone or tablet — from Anthropic's official remote control to raw SSH tunnels. Each trades off convenience against control.

# Start remote control server on desktop
claude remote-control

# Or with worktree isolation + multi-session
claude remote-control --spawn worktree --capacity 8

# Interactive: local + remote simultaneously
claude --remote-control

# Install Tailscale on both devices
curl -fsSL https://tailscale.com/install.sh | sh
tailscale up

# On desktop: start tmux with Claude Code
tmux new -s jarvis
claude

# From phone (Termux/Blink): connect
mosh user@desktop-tailscale-ip -- tmux attach -t jarvis

# Install ttyd
brew install ttyd   # macOS
apt install ttyd    # Ubuntu/Debian

# Start web terminal with Claude Code
ttyd -p 7681 -c user:pass bash -c "claude"

# Access at http://localhost:7681
# Use Tailscale or ngrok for remote access

The voice layer transforms Claude Code from a terminal tool into something that speaks. Five approaches exist, ranging from simple TTS overlays to full wake-word-activated conversational pipelines running entirely on-device.

git clone https://github.com/nickpending/clarvis
cd clarvis
bun install
# Configure as Claude Code stop hook

pip install samantha-voice-assistant

# Add as MCP server to Claude Code
claude mcp add samantha -- samantha

pip install pvporcupine pvcheetah pvorca pvrecorder pvspeaker anthropic

# Pipeline: Wake Word → STT → Claude API → TTS
# See picovoice.ai for full implementation guide

Messaging apps provide the most natural "Jarvis" interface — you text your AI the same way you text a person. Claude Code now has native Telegram and Discord support, plus community bridges for WhatsApp, Signal, Slack, iMessage, and Teams.

# 1. Create bot via Telegram @BotFather
# 2. Get bot token
# 3. In Claude Code:
/telegram:configure
# 4. Paste token when prompted
# 5. Text your bot from any device

Home Assistant is the clear winner for smart home integration with Claude. The official MCP server exposes 70+ tools — lights, thermostats, switches, automations — all accessible via natural language through Claude Code.

The key to an always-on Jarvis is daemon operation — Claude Code running headless, responding to events rather than interactive prompts. Headless mode (--print) is production-ready. A combined remote-control + headless mode is the missing piece for true cloud-native agents.

# Simple headless call
claude -p "Summarize today's unread emails"

# JSON output with tool restrictions
claude -p "Check server status" \
  --output-format json \
  --allowedTools "Bash,Read" \
  --max-turns 5

# With MCP servers and budget cap
claude -p "Turn off all lights" \
  --mcp-config ~/.mcp/ha.json \
  --max-budget-usd 0.50

MCP (Model Context Protocol) servers are the "hands and feet" of a Jarvis system. They extend Claude Code with structured access to external services — voice I/O, smart home control, browser automation, and system management.

npx @shantur/jarvis-mcp --install-claude-code-config --local

Six reference architectures, ordered by complexity. Each is a viable path to a Jarvis-like system. Most real deployments combine elements from multiple patterns.

Copy-paste commands to get each component running. Start with Path A (fastest), then layer on additional capabilities.

# On your desktop/laptop
claude remote-control

# Scan QR code with Claude app
# Done. Full session from phone.

# 1. Open Telegram, message @BotFather
# 2. Send: /newbot
# 3. Follow prompts, get token
# 4. In Claude Code:
/telegram:configure

# Paste your bot token
# Text the bot from any device

git clone https://github.com/nickpending/clarvis
cd clarvis && bun install

# Or for simpler TTS:
npx agent-vibes

# Option 1: Samantha MCP (recommended)
pip install samantha-voice-assistant
claude mcp add samantha -- samantha

# Option 2: Picovoice pipeline
pip install pvporcupine pvcheetah pvorca \
  pvrecorder pvspeaker anthropic

# Option 3: Browser-based (Jarvis MCP)
npx @shantur/jarvis-mcp \
  --install-claude-code-config --local

# In Home Assistant:
# Settings → Integrations → Add → MCP Server

# In Claude Code MCP config:
# Add your HA instance URL + token
# See: home-assistant.io/integrations/mcp_server/

# 1. Provision Hetzner VPS (~$5/mo)
# 2. Install Docker + Coolify
# 3. Deploy telegram_daemon.py
# 4. Configure:
#    - CLAUDE.local.md (brain)
#    - CRONTAB.md (schedule)
#    - memory.md (hot memory)
#    - MCP server configs
# 5. Set access control matrix

# See: okhlopkov.com/always-on-ai-agent-server-setup/