ChatGPT Remote MCP: Self-Hosted Remote Development Environment with Docker and Cloudflare

When using conversational AI models as coding assistants, one of the most frustrating bottlenecks is the model's disconnection from your actual execution environment. Copying code back and forth between your browser and editor, applying changes, running tests in a separate terminal, and pasting terminal errors back into prompts slows down your workflow.

The Model Context Protocol (MCP) was introduced to solve this exact problem. However, many existing MCP servers are designed purely as local stdio processes tied to a single desktop application on the same machine. What if you want to work remotely from a laptop, interact via ChatGPT web, or use a high-performance home PC or Linux server as a headless remote development workstation?

The open-source project chatgpt-remote-mcp is a self-hosted development server built specifically to bridge this gap. Running via Docker Compose on Windows or Linux, it combines Cloudflare Tunnel, an OAuth 2.0 authentication framework, 21 comprehensive development tools, and asynchronous process management to let ChatGPT securely interact with your machine from anywhere.

Three Real-World Challenges with Remote MCP

Building a toy HTTP MCP server for localhost is straightforward. But when you try to connect it to ChatGPT for real remote access, you quickly run into three obstacles.

First, exposing open ports to the public internet creates severe security risks. Opening router firewall ports or configuring DDNS on a private development machine invites port scans and malicious bots.

Second, ChatGPT enforces strict integration standards. OpenAI's custom MCP interface requires an OAuth 2.0 flow for secure authorization delegation. Servers lacking standardized authorization and token endpoints fail during client registration.

Third, real coding workflows require background process management. Development isn't just about reading files; you need to install packages, run builds, launch test runners, and inspect streaming stdout/stderr buffers over time. A simple one-shot execution tool cannot handle interactive workflows.

subgraph

direction

ChatGPT Request

Router Inbound Port Open

Public IP Directly Exposed

Vulnerable to Botnets & Port Scanning

end

ChatGPT Request

Cloudflare Edge (DDoS Shield & WAF)

Encrypted Outbound Tunnel

Docker Nginx & OAuth Approval Key Gate

Architecture and Core Features

chatgpt-remote-mcp is structured into a tunneling layer, a reverse proxy layer, and an Express-based MCP application layer.

1. 01subgraph
2. 02CG
3. 03end
4. 04CFT
5. 05CFD
6. 06NG
7. 07APP
8. 08FS
9. 09PM

1. Zero Inbound Port Forwarding via Cloudflare Tunnel

When the server starts, the internal cloudflared container establishes an outbound tunnel to Cloudflare's edge network. No inbound firewall ports need to be opened on your router. You get automated SSL certificates and DDoS protection out of the box simply by specifying your Cloudflare Zero Trust tunnel token in .env.

2. Built-in OAuth 2.0 Server with Approval Keys

chatgpt-remote-mcp includes a complete OAuth 2.0 authorization server implemented in Express. When ChatGPT requests client authorization via /oauth/authorize, an explicit master Approval Key is required to complete the grant.

• Approval Key Validation: Only clients presenting the generated secret key can obtain tokens, preventing unauthorized parties from connecting even if they discover your tunnel domain.
• Client Registration Capacity: Bounded by MCP_OAUTH_MAX_REGISTERED_CLIENTS (default 256) to protect memory and prune inactive clients safely.
• Token Lifecycle: 1-hour access tokens and 30-day refresh tokens provide a practical balance between security and convenience.

3. Nginx Boundary and IP Normalization

A lightweight Nginx container sits in front of the application. It normalizes client IP headers (X-Forwarded-For) at the trust boundary so that Express can enforce accurate per-client rate limits.

4. 21 Rich MCP Development Tools

The server provides 21 granular tools for interacting with files and processes:

In particular, read_files enables ChatGPT to inspect multiple source files in one turn, reducing latency and context token overhead.

5. Long-Running Process Buffer Management

Longer tasks such as running test suites or building bundles cannot be cut short by rigid HTTP timeouts. The process manager decouples processes asynchronously and buffers stdout and stderr in memory. ChatGPT can inspect incremental logs with read_process and send inputs via write_stdin.

subgraph

direction

Filesystem & Batching
list_directory, read_files

Editing & Patching
write_file, replace_in_file, apply_patch

Commands & Scripts
exec_command, run_script

end

Spawn Background Process

Stream stdout & stderr into Ring Buffer

ChatGPT inspects buffer incrementally with read_process

Inject input when needed via write_stdin

Quick Start Guide

chatgpt-remote-mcp is optimized for Windows 11 with PowerShell 7 and Docker Desktop, but runs smoothly on any Linux host with Docker Compose v2.

Step 1: Clone and Initialize Keys

Clone the repository:

Copy1git clone https://github.com/munlucky/chatgpt-remote-mcp.git2cd chatgpt-remote-mcp

Generate independent secret keys:

Copy1.\scripts\setup-keys.ps1

This generates an initial .env file with secure random secrets for the OAuth approval key and internal health probes.

Step 2: Configure Environment Variables

Edit .env to match your domain and workspace path:

Copy1# Public domain routed through Cloudflare2PUBLIC_DOMAIN=mcp.yourdomain.com3 4# Cloudflare Zero Trust Tunnel Token5CLOUDFLARE_TUNNEL_TOKEN=eyJhIjoi...6 7# Host workspace path to expose8HOST_WORKSPACE_PATH=C:\dev\my-project9 10# Container mount path11CONTAINER_WORKSPACE_PATH=/workspace

Step 3: Verify and Start

Validate Docker status, mount paths, and environment settings:

Copy1.\scripts\verify-env.ps1

If checks pass, start the server:

Copy1.\scripts\start.ps1

Step 4: Copy the Approval Key

Retrieve your OAuth approval key:

Copy1.\scripts\get-approval-key.ps1

Step 5: Connect to ChatGPT

1. In ChatGPT, open Settings and navigate to Developer or Model Context Protocol settings.
2. Click Add MCP Server.
3. Enter your tunnel URL: https://mcp.yourdomain.com/mcp
4. Choose OAuth 2.0 and provide your Approval Key in the authorization prompt.
5. Once connected, all 21 tools will be active in your chat sessions.

Practical Usage Examples

Here are common ways to use the server in conversation:

1. Codebase Exploration and Direct Patching

Copy1Prompt:2"Read package.json and src/index.ts, design a new /health route,3and apply the patch directly to the code."

ChatGPT uses read_files to collect both files in one turn, plans the route, and calls apply_patch or replace_in_file to update the codebase on your machine.

2. Running Test Suites in the Background

Copy1Prompt:2"Run npm test in the background, monitor the process output,3and diagnose any failing test cases."

ChatGPT spawns the test runner, polls the process buffer via read_process, and summarizes the test results.

3. Reviewing Usage Telemetry

Administrators can inspect aggregate tool usage over the past 24 hours:

Copy1.\scripts\usage-report.ps1 -Hours 24

This filters out automated health probes and displays tool calls and active sessions.

Summary

chatgpt-remote-mcp turns any Windows or Linux workstation into a secure, private AI development server. With Cloudflare Tunnel protecting your network perimeter and an integrated OAuth 2.0 server enforcing access control, you can give ChatGPT direct access to your codebase without sacrificing security.

Explore the open-source repository at github.com/munlucky/chatgpt-remote-mcp to set up your own remote development bridge.