In my previous post, I shared the journey that led to the creation of AiderDesk - a desktop application born from the desire to overcome the limitations of the Aider CLI. My goal was to build a smoother, more integrated AI coding experience.

But as any developer knows, the journey of improvement never truly ends. While AiderDesk significantly enhanced the code editing process, I soon found myself asking: what if AI could do more than just edit code? What if it could interact with the world around my codebase — the web, APIs, and other services?

This question led to the next significant evolution of AiderDesk: the implementation of Agent Mode and comprehensive support for Model Context Protocol (MCP) servers.

The need for more than just code

Aider is phenomenal at what it does: modifying code within a given context. But modern development is more than just writing code. It involves a rich ecosystem of tools and information sources. I realized that for an AI to be a true partner, it needed to interact with this ecosystem. It needed to be able to:

1. Browse the web for documentation or solutions to obscure errors.
2. Interact with a GitHub repository to create an issue or review a pull request.
3. Search through a Confluence or Notion workspace for project specifications.
4. Connect to thousands of apps via services like Zapier.

This desire for broader capabilities was a sentiment I saw echoed in the Aider community, with multiple pull requests and discussions about adding MCP and tool support (Aider-AI/aider#3937, Aider-AI/aider#3672). The community wants an Aider that could do more.

The gateway to a bigger world: integrating Model Context Protocol (MCP)

The solution was clear: AiderDesk needed to support the Model Context Protocol (MCP). MCP is a standard that allows an AI model to discover and use external tools. By integrating MCP, I could give AiderDesk the "hands" to interact with the outside world.

Instead of being limited to the local codebase, an MCP-enabled AiderDesk could connect to servers that provide tools for web browsing (Puppeteer MCP Server), GitHub interaction (GitHub MCP Server), or even enterprise data management (Atlassian MCP Server). This opens up a universe of possibilities and transforms the AI from a coder into a researcher, a project manager, and a systems integrator.

But this led to a new, fundamental question: if you give an AI a toolbox, how does it know which tool to use, when, and for what purpose?

The brains of the operation: the birth of agent mode

Simply supporting MCP servers wasn't enough. An orchestrator was needed - a "brain" to manage the new "hands”. This is precisely why Agent Mode was created. It was a necessary evolution driven by the integration of MCP.

Agent Mode provides an autonomous layer on top of the toolset. It allows the AI to:

1. Analyze a high-level goal from the user.
2. Plan a sequence of actions to achieve that goal.
3. Select the appropriate tool for each step, whether from an MCP server or a built-in function.
4. Execute the tool and interpret the results.
5. Iterate and adapt the plan based on the outcome.

In essence, Agent Mode gives the AI the cognitive architecture to use its tools effectively. Without it, MCP support would be a collection of disconnected features. With it, AiderDesk becomes a true agent capable of tackling complex, multi-step tasks from start to finish.

An unexpected discovery: the power of internal tools

My initial vision was to keep AiderDesk unopinionated, acting purely as a platform for users to connect their own MCP servers. I believed this would offer maximum flexibility. However, as I developed Agent Mode, I realized this approach had a flaw: it created a significant barrier to entry. A fresh installation of AiderDesk would have a powerful agent brain, but no hands to work with until the user configured external servers.

This led to the creation of built-in Power Tools.

I decided to bundle a core set of essential tools directly into AiderDesk. These tools, which handle tasks like semantic search, file system operations (read, write, edit), and running shell commands, provide a powerful out-of-the-box experience. The agent can immediately start performing useful, complex tasks without any external setup.

This hybrid approach became the new vision: AiderDesk is powerful on its own, but infinitely extensible through MCP. You can start working with the agent right away using its built-in Power Tools, and then gradually connect more specialized MCP servers as your needs evolve. Aider's own code-editing capability also became just another powerful tool in the agent's arsenal, used when the task specifically involves complex code generation or modification.

The journey continues

Creating AiderDesk was never just about fixing a few annoyances; it was about building the AI-assisted coding environment I truly wanted to use. With Agent Mode and MCP support, AiderDesk has evolved from a better UI for Aider into a true development partner. It's a testament to the idea that sometimes, the best way to improve something is to build your own version, tailored precisely to your needs and vision.

I'm incredibly excited about these new capabilities and how they empower developers to work more efficiently and autonomously.

Check out AiderDesk on GitHub: https://github.com/hotovo/aider-desk