AI Coding Workflow

AI ProgrammingClaude Coden8nAutomationWorkflow

For the past three months, I've upgraded my Claude Code workflow from "VS Code manual operations" to "DesktopCommanderMCP direct terminal control + n8n automated deployment." This is my real-world recap of the configuration steps, the pitfalls I hit, and how everything connects.

Why I Built This Pipeline

I use Claude Code daily for code refactoring and script writing. My old flow was: Claude Code generates code → I manually copy to VS Code → I manually deploy to the server. Two pain points:

1. Terminal-IDE-Deployment fragmentation: Claude Code output can't write directly to remote servers, so I had to manually scp/rsync every time

2. Repetitive operations eating time: Every code change required "generate → copy → deploy," taking 5-10 minutes per iteration

DesktopCommanderMCP solved the first problem by letting Claude Code execute terminal commands directly on the filesystem. n8n solved the second by triggering automatic deployment on file changes.

Full Architecture

Claude Code (reasoning layer)
    ↓ MCP protocol
DesktopCommanderMCP (terminal control layer)
    ↓ local filesystem read/write + diff editing
n8n (automation layer)
    ↓ Webhook / Git webhook trigger
Automated deployment script (server)

Three-layer split: Claude Code handles reasoning and generation, DesktopCommanderMCP executes terminal commands, n8n listens for changes and triggers deployment.

DesktopCommanderMCP Core Setup

DesktopCommanderMCP is an open-source MCP server that gives Claude Code terminal control capabilities. Project: wonderwhy-er/DesktopCommanderMCP.

Installation (tested on Ubuntu 24.04)

# 1. Check Node.js version (requires ≥18)
node --version  # v22.x works fine

# 2. Install globally
npm install -g @wonderwhy/desktop-commander-mcp

# 3. Verify
npx desktop-commander-mcp --version

Claude Code Configuration (cline_settings.json)

{
  "mcpServers": {
    "desktop-commander": {
      "command": "npx",
      "args": ["desktop-commander-mcp"],
      "env": {}
    }
  }
}

Once configured, Claude Code can run ls, cd, cat, diff directly in the terminal—no window switching needed.

Verify MCP Connection

Test in a Claude Code conversation:

> Run ls -la /home/projects to verify MCP connection

If it returns the directory listing, the connection works. If you get an error, check Node version and npx path.

n8n Automated Deployment Setup

With DesktopCommanderMCP handling file edits, n8n monitors Git webhooks to trigger automatic deployment.

n8n Workflow Configuration Key Points

Trigger node: GitHub Webhook / GitLab Webhook (fires on push)

Execution node sequence:

1. SSH Node: connect to server

2. Execute Command Node:

   cd /var/www/project && git pull origin main && pm2 restart app

3. Notify Node (optional): post deployment notification to Slack/ DingTalk

Pitfall 1: SSH Fingerprint Verification Failure

**Problem**: n8n SSH Node failed on first connection with Host key verification failed.

Fix:

# Manually SSH once from the machine running n8n
ssh -o StrictHostKeyChecking=accept-new user@your-server.com

n8n doesn't auto-accept new host fingerprints. Running this once writes the fingerprint to ~/.ssh/known_hosts.

Pitfall 2: pm2 Port Conflict on Restart

**Problem**: pm2 restart failed with EADDRINUSE because the old process hadn't fully exited.

**Fix**: Add wait in n8n's Execute Command Node:

pm2 stop app && sleep 3 && pm2 start app

Skill Reuse from addyosmani/agent-skills

Google Chrome engineer addyosmani's agent-skills repo (⭐74K+) provides production-grade skills for AI coding agents. After filtering, two skills are directly applicable to my pipeline:

skill 1: `git-interactive-rebase`

Auto-generates and executes rebase commands, avoiding manual merge conflict handling.

skill 2: `bash-error-diagnosis`

Captures command errors and auto-diagnoses with fix suggestions, cutting my debugging time.

Integrate these into Claude Code's custom instructions:

{
  "skills": [
    "file:///home/user/.claude/skills/git-interactive-rebase.md",
    "file:///home/user/.claude/skills/bash-error-diagnosis.md"
  ]
}

Full Pipeline Demo

Taking "modify API timeout config and auto-deploy" as an example:

1. **Claude Code**: Change timeout from 30s to 60s in config.yaml

2. **DesktopCommanderMCP**: runs sed -i 's/timeout: 30/timeout: 60/' config.yaml, local diff verification

3. **Claude Code**: Commit and push to main

4. GitHub Webhook → triggers n8n Workflow

5. **n8n** → SSH to server → git pull && pm2 restart

6. Notification: deployment success posted to Slack

Full process from "manual edit" to "AI reasoning + auto-deploy" runs automatically in ~30 seconds instead of 5-10 minutes.

Bonus: obra/superpowers Skill Framework

obra/superpowers provides a more structured skill framework for packaging DesktopCommanderMCP + agent-skills into reusable workflows. Its core idea: encapsulate "operation sequences" into atomic skills, shareable across a team.

TL;DR

The key value of this pipeline: turning AI coding from "generate-manually-deploy" fragmentation into "reason-execute-deploy" fully automated closed loop. Three tools with clear roles: Claude Code thinks, DesktopCommanderMCP controls the terminal, n8n automates deployment.

If you're already using Claude Code for development, I strongly recommend setting up DesktopCommanderMCP first—it solves a real efficiency pain. The n8n deployment part is optional: if your project only updates a few times per day, manual deployment is fine.