I recently looked into Microsoft Copilot Studio because Microsoft now supports connecting external MCP servers to it. My first impression was that Copilot Studio is very different from Claude Code, and I think the difference is mainly in the harness.

Claude Code is essentially a professional agent workbench. The LLM is surrounded by a strong development harness: it can inspect a codebase, read and modify files, run Python or shell commands, use Git, work with Excel and other files, call MCP tools, inspect the results, fix errors, and repeat the process. The agent is given something close to a computer and is expected to figure out how to get the job done.

Copilot Studio is built around a different idea. It is an enterprise agent platform designed to connect AI with Microsoft’s ecosystem: Teams, Outlook, SharePoint, Excel, Power Automate, Dataverse, and enterprise identity and governance. It can also connect to MCP servers. The agent can make judgments and select tools, but the environment is much more controlled and structured.

This makes Copilot Studio particularly good for business workflows. For example, an agent could receive an email, look up information in SharePoint, call a FactSet MCP tool, decide what type of request it is, create a task, ask a human for approval, and send the result through Teams. This is exactly the kind of AI automation that is difficult to build and maintain from scratch.

But I would not assume that connecting my FactSet MCP server to Copilot Studio suddenly gives me a Claude Code-level agent. The MCP server provides the capabilities, but the harness determines how intelligently and flexibly those capabilities are used. For complex work such as reading a large Excel universe, writing Python, running calculations, inspecting the results, changing the methodology, rerunning the analysis, and producing the final workbook, Claude Code is much closer to the environment I want.

So I see them serving different layers. Claude Code is an agent workbench for building and doing sophisticated work. Copilot Studio is an enterprise platform for putting AI into business processes. My FactSet MCP can sit underneath either one as the specialized financial intelligence layer.

The interesting architecture is therefore not necessarily to replace Claude Code with Copilot Studio. It may be to use Claude Code to build sophisticated agents and MCP services, while using Copilot Studio as the enterprise front door that connects those capabilities to Teams, Outlook, SharePoint, approvals, and other Microsoft workflows.

In short, Claude Code helps the agent do complex work; Copilot Studio helps the enterprise put AI into its existing business processes. The difference is largely the harness.