ใช้งาน ChatGPT ทำงานบนเครื่องเราฟรีไม่จำกัด | ฉบับจับมือทำตั้งแต่ศูนย์ สายฟรีใช้ได้ สายเปย์ยิ่งใช้ดี

Overview

This video is a step-by-step guide (from scratch) for using GPT chat (via Luna) together with local tools through Model Context Protocol (MCP). The setup enables GPT to read/write files on the user’s own computer using a free/unlimited GPT account workflow, without manually exposing a port to the internet.

---

Goal / Workflow

1. Set up an MCP File System server (local)

   • Use Node.js + npx to run a local MCP server.

2. Verify the server using MCP Inspector

   • Confirm the server is reachable and that the expected tools are available.

3. Connect GPT chat to the local MCP server

   • Use an OpenAI “tunnel” so you typically don’t need to open PC ports yourself.

4. Create a GPT chat plugin

   • Build a “Local Workspace” plugin that routes GPT tool calls to the tunnel/MCP server.

---

Free vs. Paid Accounts (As Claimed)

• The speaker claims the same technique applies to both free and paid GPT accounts.
• With a paid setup, additional features may be available, but it still uses the MCP + tunnel approach.

---

MCP File System Server Setup

Prerequisites

• Install Node.js (the video references version checks such as Node 26 and mentions NPX usage).

Create a demo workspace

• Create a working directory and a demo folder.
• Add an example file such as Hello.txt.

Start the MCP File System server

• Run the MCP server via npx, described as running on STDIO, for example:

npx -y @modelcontextprotocol/server-filesystem

Confirm with MCP Inspector

• Check that:
  • The connection succeeds (e.g., green status indicators).
  • Tools exist, including read-like capabilities (e.g., Read Text File, Read Media), and appropriate write-like abilities.
• Ensure the server is restricted to the intended folder scope—permission/scope boundaries matter.

---

Why a Tunnel Is Needed

• GPT runs in the cloud, while MCP runs locally.
• Without a tunnel, you would likely need to open a local port (a security concern).
• The video recommends using OpenAI’s tunnel, described as a mechanism built specifically for GPT chat to reach MCP securely.

---

Tunnel Setup + Credentials (Security Emphasis)

Enable Developer Mode

• Turn on Developer Mode in the GPT/Tunnel UI to create plugins/tunnels.

Create a tunnel and collect credentials

You receive:

• Tunnel/Channel ID
  • Used like a “house number.”
• Control Plane API Key / secret key
  • Used like a “house key.”

Warning: API keys must be kept secret. The speaker suggests deleting them after use.

---

Local Tunnel Client Setup

Run the tunnel client

• Download and run the tunnel client executable (e.g., something like tunnelclient.exe).

Use a YAML configuration file

• The speaker notes YAML isn’t overly complex.
• The config includes:
  • Tunnel ID
  • API key reference
  • Which MCP server command to start (so the tunnel knows the backend to launch)
  • A port selection (the video mentions a default around 8080, but allows changes and discusses occasional environment/connection quirks)

---

End-to-End Testing (Smoke Test)

After starting the tunnel client and binding it to the MCP server:

• In GPT chat (via the “Local Workspace” plugin), ask something simple like:
  • “What’s inside Hello.txt?”
• GPT should return the file contents, proving the tool calls reached the local MCP server.

---

Switching MCP Backends (Extensibility Concept)

A major theme is that you can swap the MCP backend without changing the overall “house/tunnel/plugin” structure.

• Example backend replacement:
  • Start with file system MCP
  • Then switch to Serena MCP
  • Later switch to STI MCP for website/design generation
• The speaker frames this as assembling different “robots” by combining:
  • tunnels + different MCP servers

---

Automating Website Design + Export

Later in the demo, the workflow described is:

1. Generate/modify a website prototype
2. Fetch resulting files back to the local workspace
3. Hand off to Serena to convert the frontend into a production-ready output
   • Described as converting HTML prototype → React → Next.js / final website steps

The intent is to reduce manual steps like exporting and continued polling.

---

Multiple Plugins / Multiple MCP Servers

• The video notes issues when trying to use multiple tunnels / multiple “plugin backends” simultaneously with a single plugin.
• Suggested workaround:
  • Open another terminal
  • Start another MCP server
  • Create another tunnel/plugin if needed
  • Coordinate downloads and subsequent processing

---

Practical Tips / Troubleshooting

• Starter kit bundles may fail due to environment differences.
  • Recommendation: have the AI debug and fix errors by asking it to resolve issues.
• Common troubleshooting examples:
  • Mismatched config IDs
  • Errors in tunnel/channel ID formatting (e.g., a missing digit)

---

Guides / Tutorial Elements Highlighted

• “From zero” setup
• Step-by-step commands:
  • Node.js install
  • npx command usage
  • Running MCP server
• Inspector-based verification
• Tunnel + plugin configuration
• Final smoke test (reading Hello.txt)
• Backend swapping (filesystem MCP → Serena MCP → STI MCP)
• Automation concept for multi-step website generation and conversion

---

Main Sources / Components Mentioned

• Main speaker/instructor: a single creator/instructor (no clear name given in subtitles)
• Core components:
  • OpenAI GPT chat / “Luna”
  • Model Context Protocol (MCP)
  • MCP Inspector
  • OpenAI Tunnel
  • Local Node.js/npx tooling
• Backends referenced:
  • Serena
  • STI