Coding Tools MCP: Turning ChatGPT into a Real Coding Agent

ChatGPT Use cases and examples
1

Hi everyone,

I’ve been experimenting with MCP-based workflows recently, and one thing kept bothering me:

ChatGPT is incredibly capable at writing code, but it normally can’t interact with a real project.

It can’t inspect a repository, search files, run commands, check Git status, or apply changes directly. Most of the workflow still involves copying and pasting context back and forth.

To address this, I built Coding Tools MCP, an MCP server that exposes practical coding operations to ChatGPT and other MCP-compatible clients.

What it provides

Coding Tools MCP exposes a set of tools commonly needed during software development:

  • Read files and browse directories

  • Search across a codebase

  • Apply patches to files

  • Execute terminal commands

  • Interact with long-running processes

  • Inspect Git repositories

  • Request permission escalation when needed

  • View local images

The goal is simple:

Give AI assistants direct, controlled access to a real development workspace.

Example workflow

Instead of:

“Here’s a file. Can you review it?”

You can ask:

“Analyze this repository and explain its architecture.”

Instead of:

“Here’s the error log.”

You can ask:

“Find the root cause of this bug and propose a fix.”

Instead of:

“Generate a patch.”

You can ask:

“Implement this feature and update the relevant files.”

The model can then use the MCP tools to inspect the project, search code, run commands, and modify files as needed.

Security

The project was designed around workspace confinement and explicit permissions.

It supports:

  • Workspace-scoped file access

  • Permission requests for sensitive operations

  • Runtime policy controls

  • Git-aware tooling

  • Process lifecycle management

The intention is to make powerful coding capabilities available while keeping operations transparent and auditable.

Why I built it

I wanted something that felt closer to a coding agent than a traditional chatbot, while remaining compatible with the MCP ecosystem and easy to self-host.

With ChatGPT Apps and MCP support becoming more widely available, it felt like the right time to build a dedicated coding-focused MCP server.

Project

GitHub: xyTom/coding-tools-mcp

I’d love to hear feedback, feature requests, or ideas from others building MCP-powered developer tools.

2

I think you just made a bad version of codex ^^’

As far as I know, Codex can do all of that :

“It can’t inspect a repository, search files, run commands, check Git status, or apply changes directly. Most of the workflow still involves copying and pasting context back and forth.”

If GPT is the brain codex is the brawl.

3

For me personally working in codex even from my phone and connected to my computer this way easier than working in chat.

Then I understand the logic why you would like to have chat access to your project (lake planning and other things). Personally, I would not allow chat the right code directly in my project because better access vehicle codecs sounds more solid to me for various reasons. What what might be useful is the ability of chart to create files somewhere in the documentation of the project or its planning or some research sources and other things.

There would be basically two ways to connect them. I think the easiest one would be the github MCP server which you would connect to chat and allow chat to read and create files in your project on github. This is probably the easiest setup. I didn’t find the official connector working correctly for me maybe I just don’t know how to set it up, but I had a couple of projects where I added custom MCP server to a custom chat GPT and that MCP server was the github MCP server. And that was a good enough.

Is the other way to connect them would be little bit more tricky: set up a secure tunnel via cloudflare and connect chat GPT to local IDE MCP server. That would be awesome for things like code reviews codebase research and other things. Not very complicated to create, budget would require you to start your IDE manually and the tunnel each time you want to work. So it is way easier just to set up Codex and work from your phone.

4

Nice direction. The thing I would make very explicit in the server docs is the separation between inspection, planning, and mutation tools.

For a coding MCP, I would want the tool surface to make these boundaries visible before any call runs:

  • read/search/list operations are always safe to call
  • write/apply/run commands require an explicit working directory and scope
  • destructive operations require dry-run output first
  • every command returns enough diagnostic context for the model to recover without guessing
  • the server logs which files, commands, and git state were touched

The MCP part makes the workflow possible, but the trust comes from the contract around side effects.