> ## Content Index
> Fetch the complete content index at: https://www.anasabaziz.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# How I Upgraded From an AI Chatbot to a Full AI Agent Coding Platform
- URL: https://www.anasabaziz.com/how-i-upgraded-from-an-ai-chatbot-to-a-full-ai-agent-coding-platform/
- Published: 2026-08-29T01:51:26.000Z
- Updated: 2026-08-29T01:51:26.000Z
- Description: Learn how to connect AI agents to apps like PowerBI or GitHub, from setting up projects and skills to managing tokens, subscriptions and APIs.
- Author: Anas
- Tags: LLM

Have you seen people connecting their AI chats to tools like Power BI or GitHub and wondered how they actually do it? The good news is that you don't need to be an AI engineer to start experimenting with this setup.

In this guide, I'll walk through the approach I use with the OpenAI (ChatGPT) ecosystem, from installing the desktop app and organising project files to connecting an MCP server and eventually moving beyond subscription-based usage limits.

TL;DR / Quick Verdict click to expand ▾ 

- Start with a ChatGPT account and desktop app before worrying about advanced AI infrastructure.
- Keep each AI project in its own folder so documents, skills and instructions don't get mixed together.
- Use Agent Skills to define how the agent should work, while MCP provides the connection to external tools and data.
- For Power BI and Fabric, Microsoft's skills-for-fabric repository is a useful example because it brings together skills and MCP-related information.
- Free or lower-tier subscriptions are useful for experimentation, but larger agent workflows can consume tokens quickly.
- If subscription limits become the bottleneck, consider a higher tier, local open-weight models, or an AI aggregator such as OpenRouter.

---

**Table of Contents** 
- [What Are We Actually Connecting?](#section-1)
- [Step 1: Start With a ChatGPT Account](#section-2)
- [Step 2: Install the ChatGPT Desktop App](#section-3)
- [Step 3: Create a Dedicated Project Folder](#section-4)
- [Step 4: Connect an MCP Server](#section-5)
- [Step 5: Understand Token Usage and Subscription Limits](#section-6)
- [Step 6 and Beyond: Move to Local Models or Pay-as-You-Go](#section-7)
- [Open-Weight Models vs AI Aggregators](#section-8)
- [Final Verdict: Start Simple, Then Add Complexity](#section-9)

## What Are We Actually Connecting?

When people talk about connecting an AI chat to Power BI, GitHub or another piece of software, there are several moving parts underneath the surface. The important distinction is between the AI model itself, the instructions or capabilities you give the AI agent, and the external systems it can interact with.

For this guide, I'll use the OpenAI (ChatGPT) ecosystem as the example. That's mainly because it's a practical comparison for people who already encounter Microsoft Copilot in business environments that defaults to using OpenAI frontier models. The exact approach will differ between products, but competitors such as Claude, Gemini and Grok have similar concepts around giving AI access to tools and external context.

The key technology we'll focus on is the Model Context Protocol, or MCP. Think of an MCP server as a lightweight bridge between an AI application and external tools, data sources or workflows.

Installing an AI application doesn't automatically mean it can understand or safely manipulate everything on your computer. You need to deliberately give the agent the right instructions, files and connections.

## Step 1: Start With a ChatGPT Account

The first step is straightforward: you'll need a ChatGPT account. Sign up at [ChatGPT.com](https://chatgpt.com/?ref=anasabaziz.com) if you havent done so.

If you're only trying this out, you don't necessarily need to start with the most expensive subscription. A FREE or GO subscription can be enough to begin experimenting with the workflow and understand what AI agents can actually do.

I wouldn't recommend paying for a higher tier before you've established that this type of workflow is useful to you. The better approach is to start small, connect one project, and see how much interaction you actually need.

The important thing is to think beyond the normal chat interface. We're moving toward a workflow where the AI isn't simply answering a question from information you paste into a chat. Instead, you're giving it a structured project environment and, eventually, controlled access to external tools.

## Step 2: Install the ChatGPT Desktop App

Next, install the ChatGPT desktop application. Previously there was a separate Codex app. But now ChatGPT Chat, Codex and Work is bundled into one desktop app.

The application is available on Windows, MacOS & Linux (*preview as at August 2026*). Don't worry about needing a specific Operating System to run it. Simplest way is to download from the [Codex website](https://openai.com/codex/?ref=anasabaziz.com), Windows allows you to download from the Microsoft Store too.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-3-1.png)

ChatGPT desktop app available on Microsoft Store for Windows

Once the application is installed, you can start exploring the agent-style workflow. This is where the experience becomes quite different from simply opening a browser and asking ChatGPT a question.

The biggest difference is now you can let ChatGPT control apps on your computer. Ensure this option is turned on. If it isn't, it's still just an AI chatbot. 

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-9-1.png)

Computer use options in the ChatGPT App

One practical constraint appears quickly, though: usage limits. A FREE or GO subscription can give you enough capacity to try the workflow, but you shouldn't assume that it's sufficient for a large project. Agents can perform many interactions while working through files, instructions and external tools, and those interactions add up.

My recommendation is therefore to use the initial subscription level as a testing environment. Build something small first. Once you understand the workflow and where the actual bottlenecks are, then decide whether you need more capacity. We'll get back to this again in the further sections of this article.

## Step 3: Create a Dedicated Project Folder

This is one of the simplest steps, but it's also one of the easiest to overlook.

For each project, I recommend creating a dedicated project folder rather than throwing every document, instruction and supporting file into one general-purpose directory.

The reason is straightforward: context matters. You don't want to accidently jumble up a few projects together.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-2-1-1.png)

Codex home page

If you're working on a Power BI project, for example, you don't want the agent confusing those files with documents from an unrelated automation project. A clean project structure makes it much easier to control what the agent sees and works with.

This is also where Agent Skills become useful. You can manually place the skills you want into the relevant agents folder, or instruct the AI to include them where appropriate.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-7-1.png)

Codex Project File

I've also made some of the skills I use available through my GitHub repository, so you can use those as a starting point rather than building everything from scratch. You can find them here: 

[GitHub - anasabaziz/AI-Skills: This Repositary contains useful skills made by Anas Ab AzizThis Repositary contains useful skills made by Anas Ab Aziz - anasabaziz/AI-Skills![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/icon/favicon-7397b76d-ccd3-4782-aa91-334aa2c129f2.png)GitHubanasabaziz![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/thumbnail/AI-Skills-21254143-0a6a-43ab-b612-419d1c1766e2)](https://github.com/anasabaziz/AI-Skills?ref=anasabaziz.com)

## Step 4: Connect an MCP Server

Now we get to the part that usually makes people curious: connecting the AI agent to an external tool.

Model Context Protocol, or MCP, provides a standardized way for AI applications and large language models to interact with external data sources, tools and workflows. In practical terms, an MCP server can act as the bridge between your AI agent and something such as Power BI or Microsoft Fabric.

A useful example is Microsoft's `skills-for-fabric` repository:

[GitHub - microsoft/skills-for-fabric: A collection of skills and MCP systems to enable users of CLI, VSCode, Claude to operate over Microsoft FabricA collection of skills and MCP systems to enable users of CLI, VSCode, Claude to operate over Microsoft Fabric - microsoft/skills-for-fabric![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/icon/favicon-32f3555e-bfa7-4800-9693-458e6c94eb20.png)GitHubmicrosoft![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/thumbnail/skills-for-fabric-1e06203f-e8bc-4de2-8458-9d045fdd2f9e)](https://github.com/microsoft/skills-for-fabric?ref=anasabaziz.com)

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-5-1.png)

Prompting Codex to install the Skills-For-Fabric bundle by itself

What makes this particularly interesting is that the repository brings together information around both skills and MCP connectivity for Power BI and Fabric.

This gives you a useful mental model:

- **Skills** help define how the agent should perform a particular type of work.
- **MCP** gives the agent a mechanism for interacting with an external system.
- **Your project files** provide the context specific to what you're actually working on.

One of the first tests I'd recommend is deliberately conservative. Rather than immediately telling the agent to change your Power BI project, have it review your PBIP files first.

That review step gives you an opportunity to see whether the agent understands the structure of the project before you allow it to make amendments. It's a much better starting point than jumping straight into automated changes.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-8-1.png)

Prompting Codex to review the Power BI file

## Step 5: Understand Token Usage and Subscription Limits

At this point, you may notice something that isn't obvious when you're just using an AI chatbot for occasional questions: agent workflows can consume a lot more usage.

Every time the agent reads files, reasons about them, interacts with an MCP server and performs another step, you're potentially consuming additional model capacity.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-6-1.png)

Monthly usage limit shown in the user settings

This is why I wouldn't expect FREE or GO subscriptions to remain comfortable once you move into larger projects.

If you're regularly working with Power BI files, repositories, multiple skills and external tools, I'd look at a higher subscription tier such as PLUS or PRO.

There is an important caveat, though: higher tiers don't mean unlimited usage. You still have usage limits, although they're substantially higher. For PLUS subscriptions in particular, usage is also subject to a five-hour usage window.

So the decision shouldn't simply be "Which subscription is the most powerful?"

The better question is:   
**How much agent usage does my workflow actually require?**

If you're experimenting, start lower. If you're consistently hitting limits because you're doing meaningful work, then upgrade.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-4-1.png)

Upgrade plans available for ChatGPT

## Step 6 and Beyond: Move to Local Models or Pay-as-You-Go

Once you've reached this stage, there are two more advanced directions you can take.

The first is to run open-weight models yourself.   
The second is to use an AI aggregator and pay for model usage directly.

Both approaches can remove some of the usage-limit constraints associated with subscription plans, but there's a trade-off that shouldn't be ignored: you're also removing the ceiling on what you can spend.

That's particularly important with agent workflows. If an agent is allowed to perform many tool calls and model interactions, a pay-as-you-go setup can consume more money simply because the workflow is being used more heavily. Similarly a local model consumes more power and increases your electricity bill.

So don't think of this as a magical way to make AI cheaper. Think of it as changing the pricing model from a subscription with usage limits to direct consumption.

That can be a good trade when you need flexibility, but you need to manage the cost deliberately.

## Open-Weight Models vs AI Aggregators

If you want to run everything **locally**, tools such as [Ollama](https://ollama.com/?ref=anasabaziz.com) or [LMStudio](https://lmstudio.ai/?ref=anasabaziz.com) provide one route into open-weight models.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-12-1.png)

Open-weight models available for download on Ollama

The major advantage is control. You can run models on your own hardware rather than depending entirely on a hosted subscription or API provider. Though these companies still provide that option too.

The catch is hardware.

You need a sufficiently capable PC to run the models you want at a useful speed, and getting a good local AI setup running is a separate technical project in itself. I wouldn't recommend going down this route simply because you've hit a ChatGPT usage limit. Expect to incur thousands of dollars or perhaps tens of thousands for this option.

The alternative is an **AI aggregator**.

This is the route I personally find more practical if you already have a PC capable of running the desktop AI workflow. I use services such as [OpenRouter](https://openrouter.ai/?ref=anasabaziz.com) to purchase model tokens for consumption, then connect the application through an API key.

The process is conceptually simple: load funds, configure your API connection, and set spending limits so you don't unexpectedly burn through your balance.

Another advantage is model choice. With an aggregator, you're not restricted to ChatGPT models. You can experiment with models from different providers and see which performs best for the work you're actually doing.

For coding work specifically, I personally like Kimi K2.6 (Moonshot AI, released April 2026). It had helped me code my websites without much problems.

The important point isn't that one model is universally better. It's that this setup gives you the flexibility to choose the model based on the task rather than tying your entire workflow to one ecosystem.

![](https://storage.ghost.io/c/10/2b/102b44ac-b001-4e5c-bf76-dad05157f35d/content/images/2026/08/image-10-1.png)

Consumption activity data on OpenRouter

## Final Verdict: Start Simple, Then Add Complexity

If you're new to AI agents, I wouldn't start by building a complicated local AI infrastructure with multiple models and custom servers.

Start with the simplest version of the workflow:

**ChatGPT account → desktop app → dedicated project folder → skills → MCP server → controlled test.**

Once you've successfully connected an external system such as Power BI or Fabric and understand what the agent can actually do, you'll have much better information for deciding whether you need a higher subscription or a pay-as-you-go setup.

The other important lesson is that MCP isn't the entire solution. The quality of the workflow depends on the combination of your project structure, skills, model, external connections and the instructions you give the agent.

For something like Power BI, I'd especially recommend starting with read/review operations before allowing automated amendments. Let the agent demonstrate that it understands your files first.

---

*If you're experimenting with this workflow yourself, I'd be interested to hear what you're connecting your AI agent to. Comment below on what you're using it for*