Skip to content

Install ClawMux — Linux Split

Hub runs locally on a Linux machine (no GPU required). TTS and STT run on a remote GPU server.

Requirements: Linux (Ubuntu 20.04+/Debian 11+), Python 3.10+, Claude Code, remote Linux server with NVIDIA GPU accessible via network

1. System Check (Local Machine)

# OS
uname -s  # Must be Linux

# Python
python3 --version  # Must be 3.10+

# System dependencies (Debian/Ubuntu)
sudo apt install -y python3-venv tmux

# Claude Code
claude --version  # Must be installed and authenticated

2. Clone Repository

if [ ! -d "$HOME/GIT/clawmux" ]; then
    mkdir -p "$HOME/GIT"
    git clone https://github.com/zeulewan/clawmux.git "$HOME/GIT/clawmux"
fi
cd "$HOME/GIT/clawmux"

3. Python Environment

cd "$HOME/GIT/clawmux"
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

4. GPU Server Setup

SSH into your GPU server, clone the repo, and set up TTS/STT:

# On the GPU server:
git clone https://github.com/zeulewan/clawmux.git ~/clawmux
cd ~/clawmux

# Install Whisper and Kokoro services
bash services/whisper/install.sh
bash services/whisper/start.sh

bash services/kokoro/install.sh
bash services/kokoro/start.sh

Environment variables (optional — shown with defaults):

Variable Default Description
CLAWMUX_HOME ~/.clawmux Base directory
CLAWMUX_WHISPER_PORT 2022 Whisper STT port
CLAWMUX_KOKORO_PORT 8880 Kokoro TTS port
CLAWMUX_WHISPER_MODEL large-v3 Whisper model size

Make the services accessible from your local machine. Options:

  • Tailscale (recommended): Both machines on the same tailnet. Use Tailscale IPs.
    # On GPU server — expose via Tailscale serve
    sudo tailscale serve --bg --https=8881 http://127.0.0.1:8880
    sudo tailscale serve --bg --https=2023 http://127.0.0.1:2022
    
  • SSH tunnel: Forward ports from local to GPU server.
    # On local machine — tunnel TTS and STT
    ssh -NL 8880:127.0.0.1:8880 -L 2022:127.0.0.1:2022 user@gpu-server &
    

5. Configure Remote TTS/STT

After the hub starts, configure remote TTS/STT URLs:

# Replace GPU_SERVER with the Tailscale IP or hostname of your GPU server
GPU_SERVER="100.x.x.x"  # Or hostname.ts.net

curl -X PUT http://localhost:3460/api/settings \
  -H "Content-Type: application/json" \
  -d "{\"tts_url\": \"http://${GPU_SERVER}:8880\", \"stt_url\": \"http://${GPU_SERVER}:2022\"}"

If using SSH tunnels, the URLs are http://127.0.0.1:8880 and http://127.0.0.1:2022.

6. Register MCP Server

INSTALL_DIR="$HOME/GIT/clawmux"
claude mcp add -s user clawmux -- "$INSTALL_DIR/.venv/bin/python" "$INSTALL_DIR/server/mcp_server.py"

7. Install Slash Commands

mkdir -p ~/.claude/commands
cp "$HOME/GIT/clawmux/.claude/commands/clawmux.md" ~/.claude/commands/clawmux.md

8. Install CLI

# Review the script first
cat "$HOME/GIT/clawmux/clawmux" | head -20

# Option A: User-local install (no sudo needed)
mkdir -p ~/.local/bin
cp "$HOME/GIT/clawmux/clawmux" ~/.local/bin/clawmux
chmod +x ~/.local/bin/clawmux
# Ensure ~/.local/bin is in your PATH

# Option B: System-wide install
sudo cp "$HOME/GIT/clawmux/clawmux" /usr/local/bin/clawmux
sudo chmod +x /usr/local/bin/clawmux

9. Start the Hub

cd "$HOME/GIT/clawmux"
./start-hub.sh

Or in tmux:

tmux new-session -d -s clawmux "cd $HOME/GIT/clawmux && ./start-hub.sh"

10. Tailscale HTTPS (Optional)

For remote access from phone/tablet:

sudo tailscale serve --bg --https=3460 http://127.0.0.1:3460

11. Verify

# Hub running
curl -s http://localhost:3460/api/sessions | python3 -c "import sys,json; print(len(json.load(sys.stdin)), 'sessions')"

# TTS working (via configured URL)
curl -s http://${GPU_SERVER}:8880/v1/models | head -c 100

# STT working (via configured URL)
curl -s http://${GPU_SERVER}:2022/v1/models | head -c 100

# Browser UI
echo "Open http://localhost:3460 in your browser"

Troubleshooting

Problem Fix
TTS/STT connection refused Check GPU server services: curl -s http://127.0.0.1:2022/v1/models / curl -s http://127.0.0.1:8880/v1/models
Tailscale not connecting Ensure both machines are on the same tailnet: tailscale status
SSH tunnel dies Use autossh for persistent tunnels: sudo apt install autossh
High latency Check Tailscale direct vs relayed: tailscale ping gpu-server
MCP tools not found Wait 10s after starting Claude Code, then retry