AI hosts for Power BI Desktop: a comparison


This article compares the AI clients that connect to Power BI Desktop through the Power BI Modeling MCP server on Windows, to choose one before starting a setup.

MCP stands for Model Context Protocol. The Power BI Modeling MCP server is the same npm package in every row of the tables below, @microsoft/powerbi-modeling-mcp, started with npx as a local process. What changes between the clients is how you declare it, how you approve what it does, and whether the client can also run the Power BI report authoring skill, which works on the report of a PBIP project. PBIP stands for Power BI Project.

All ten clients support a local MCP server on Windows. Node.js is a requirement in every row, because the server is started with npx.

Installation and account

Client Microsoft Store Other installation on Windows Account
GitHub Copilot CLI no npm install -g @github/copilot, or winget install GitHub.Copilot GitHub, with a plan that includes MCP
Visual Studio Code yes download from code.visualstudio.com, or winget install Microsoft.VisualStudioCode GitHub, with a plan that includes agent mode and MCP
ChatGPT desktop app yes, identifier 9PLM9XGG6VKS download from chatgpt.com ChatGPT, paid plan
Codex CLI no installer of OpenAI, winget install -e --id OpenAI.Codex, or npm install -g @openai/codex ChatGPT
Claude Code Desktop no download from claude.com, or winget install -e --id Anthropic.Claude Claude, Pro or higher
Claude Code no installer of Anthropic, winget install Anthropic.ClaudeCode, or npm install -g @anthropic-ai/claude-code Claude, Pro or higher
Google Antigravity no download from antigravity.google/download personal Google account, free with weekly rate limits
Grok Build no installer of SpaceXAI, or winget install xAI.GrokBuild SpaceXAI, plan that includes Grok Build
Kiro no download from kiro.dev GitHub, Google, or AWS, free plan included
Qwen Code no installer of the project, or npm install -g @qwen-code/qwen-code@latest paid key of a supported provider

Two of the ten are published in the Microsoft Store. Six of them are on WinGet, and four are npm packages, so a machine without administrator rights installs most of them in the user profile.

Registration of the MCP server

Client Command Configuration file Format
GitHub Copilot CLI copilot mcp add %USERPROFILE%\.copilot\mcp-config.json JSON
Visual Studio Code the extension registers the server .vscode\mcp.json, or the file opened by MCP: Open User Configuration JSON
ChatGPT desktop app Settings, MCP servers, Add server %USERPROFILE%\.codex\config.toml TOML
Codex CLI codex mcp add %USERPROFILE%\.codex\config.toml TOML
Claude Code Desktop the Connectors panel %USERPROFILE%\.claude.json JSON
Claude Code claude mcp add %USERPROFILE%\.claude.json JSON
Google Antigravity the MCP Servers panel %USERPROFILE%\.gemini\config\mcp_config.json JSON
Grok Build grok mcp add %USERPROFILE%\.grok\config.toml TOML
Kiro the MCP Servers panel %USERPROFILE%\.kiro\settings\mcp.json JSON
Qwen Code qwen mcp add %USERPROFILE%\.qwen\settings.json JSON

Every client also reads a configuration of the project, in the folder you open, with a name that changes between them. The user scope is the simplest choice for a first setup, because it does not depend on the folder.

One difference is worth knowing before a long debugging session: Google Antigravity does not inherit the path of your terminal, because it is a desktop application, so a server declared with a bare npx can fail to start without a message that explains it. Grok Build goes in the opposite direction and also reads the MCP servers declared for Claude Code, so a server can be active without appearing in its own file.

The report authoring skill

The report layer requires the powerbi-authoring plugin, published by Microsoft in the skills-for-fabric repository. Two clients install it with a command, the others need a manual copy of the skill folders.

Client Installation of the skill Documented by Microsoft
GitHub Copilot CLI /plugin marketplace add and /plugin install yes
Visual Studio Code not available today the limit is documented
Claude Code /plugin marketplace add and /plugin install yes
Claude Code Desktop the marketplace declared in the settings, then the plugin panel partially
Grok Build the marketplace of Claude Code, read without any setting no
ChatGPT desktop app import of the marketplace by an administrator, or copy in .agents\skills no
Codex CLI clone of the repository, and copy in .agents\skills the clone, yes
Google Antigravity copy in .gemini\config\skills, or in .agents\skills no
Kiro copy in .kiro\skills no
Qwen Code copy in .qwen\skills no

The skills use the same SKILL.md format in every client, which is why the manual copy works at all. The rows marked as not documented are combinations that nobody tests: they can stop working with any update of the repository.

Two tools complete the report layer in every client, and they are installed with npm:

npm install -g @microsoft/powerbi-report-authoring-cli@latest @microsoft/powerbi-desktop-bridge-cli@latest

How to choose

  • The shortest path to the report layer is GitHub Copilot CLI or Claude Code, because the plugin installs with two commands and Microsoft documents them.
  • The shortest path to the semantic model only is any of the ten. Visual Studio Code is the only one where the report authoring skill is not offered, and the prompt that connects to Power BI Desktop is the same everywhere.
  • On a managed machine without administrator rights, prefer a client that installs in the user profile: the npm packages, the installers of Anthropic, OpenAI, and SpaceXAI, and the Microsoft Store version of the ChatGPT desktop app.
  • If you already use Claude Code, Grok Build reuses its configuration, including the marketplaces and the MCP servers.
  • If you want to see the state of the connection, Google Antigravity, Kiro, and the Claude desktop application show it in a panel, the command-line clients show it with a command.

Limitations

  • The MCP server operates on the semantic model, not on the report.
  • The report authoring skill operates on PBIP projects in PBIR format, not on a .pbix file. PBIR stands for Power BI enhanced report format.
  • Model metadata does not stay local: table names, column names, measure definitions, and query results are sent to the language model of the service you signed in to.
  • The MCP server, the report authoring skill, and the Power BI Desktop Bridge are all in preview. Behavior and tools can change before general availability.
  • The plans and the free quotas of these services change often. The tables report what the vendors published, and the linked pages are the authority.

References

The content of this article was verified in September 2026. Command names, installation channels, and plan boundaries change often in this area, so the linked documentation is the authority.

Last update: Sep 13, 2026