Remote Server Control via SSH

Remote Server Control via SSH | Generated by AI

Home 2026.04

Question: Is it a good practice to use Claude Code on a separate laptop to control a LAN server running OpenClaw (AI agent), since controlling it remotely via SSH avoids interruption issues that would occur if the agent tried to restart itself?

Answer:

Yes, this is an excellent architectural pattern and a well-established best practice in systems administration and AI agent deployment. Here’s why it works so well:

The Core Problem You’re Avoiding

When an AI agent tries to manage its own process, you get a self-referential problem:

Agent receives instruction: “restart the server”
Agent begins executing restart command
Agent’s own process dies mid-execution
No confirmation, no error handling, no recovery

This is analogous to why you don’t perform surgery on yourself.

Your Pattern: External Controller via SSH

[Your Laptop - Claude Code]
        |
        | SSH
        v
[LAN Server 192.168.1.36]
   ├── llama-server (Gemma model)
   ├── OpenClaw agent
   └── config files (openclaw.json, etc.)

Claude Code on your laptop becomes the out-of-band controller — it sits entirely outside the system it’s managing.

Why This Is The Right Approach

1. No self-interruption risk Claude Code keeps running on your laptop regardless of what happens on the LAN server. It can restart llama-server, kill OpenClaw, edit configs, and verify the restart — all without losing its own execution context.

2. Full observability Claude Code can tail logs, check process status, and confirm services came back up:

ssh lzw@192.168.1.36 "tail -f /home/lzw/llama-server.log"

3. Atomic config changes + restart

ssh lzw@192.168.1.36 "
  # Edit config
  sed -i 's/reasoning-format.*/reasoning-format none/' openclaw.json
  
  # Restart service
  kill \$(pgrep llama-server)
  nohup /home/lzw/projects/llama.cpp/build/bin/llama-server ... &
  
  # Verify
  sleep 3 && curl http://localhost:8081/health
"

4. Separation of concerns

Claude Code = management plane (your laptop)
OpenClaw + llama-server = data plane (LAN server)

This mirrors how production infrastructure works — your CI/CD pipeline (GitHub Actions, etc.) is never running on the server it’s deploying to.

Practical Tips for This Setup

Use nohup or screen/tmux on the server so processes survive SSH session drops
SSH key auth so Claude Code can connect without password prompts
Health check endpoints — Claude Code can poll http://192.168.1.36:8081/health to confirm restarts succeeded
Idempotent scripts — write restart scripts that work whether the process is running or not (pkill -f llama-server || true)

Immediate Next Step

Yes, have Claude Code SSH in and restart with --reasoning-format none:

ssh lzw@192.168.1.36 "kill 109829; nohup /home/lzw/projects/llama.cpp/build/bin/llama-server \
  -m /home/lzw/projects/llama.cpp/models/gemma-4-26B-A4B-it-UD-IQ3_S.gguf \
  --port 8081 -c 32768 --reasoning-format none \
  >> /home/lzw/llama-server.log 2>&1 &"

Claude Code on your laptop will wait for the SSH command to complete and can immediately verify the fix — something OpenClaw could never safely do for itself.

Back Donate