power-bi-agent

End-to-end Power BI development assistant for GitHub Copilot. Bundles a chat-mode agent and a set of topic skills (semantic modeling, DAX, performance, visualization, SQL→M) that work together against live models and Power BI Project files:

• powerbi-modeling-mcp — read/write the live Power BI Desktop / Fabric model.
• microsoft-learn MCP — ground every recommendation in current Microsoft Learn documentation.
• Power BI Project files (PBIP/PBIR/TMDL) — edit supported report and semantic model metadata directly when project files are present in the workspace.

What's inside

Path	What it is
agents/power-bi.agent.md	Consolidated VS Code chat-mode agent that routes to the right skill
skills/powerbi-modeling/	Star schema, relationships, RLS, calculation groups, naming, documentation
skills/powerbi-dax/	DAX authoring + an "optimize this DAX" mode with diagnosis → rewrite → why → verify
skills/powerbi-performance/	Diagnosis decision tree, model/report/refresh/capacity tuning, KQL monitoring
skills/powerbi-visualization/	Chart selection, layout patterns, accessibility, mobile design
skills/powerbi-sql/	Wrap a .sql file in an M Sql.Database() expression
plugin.json	GitHub Copilot CLI plugin manifest (root, per spec)
.mcp.json	MCP server bundle installed with the plugin
mcp/mcp.linux.json, mcp/mcp.wsl.json	Reference MCP server configs (default and WSL flavor)

Install

As a VS Code chat agent + skills (manual)

VS Code reads chat agents from ~/.copilot/agents/ and skills from ~/.copilot/skills/. Symlink (or cp -r) this repo's contents:

mkdir -p ~/.copilot/agents ~/.copilot/skills
ln -s "$(pwd)/agents/power-bi.agent.md" ~/.copilot/agents/power-bi.agent.md
for s in skills/*/; do
  ln -s "$(pwd)/$s" "$HOME/.copilot/skills/$(basename "$s")"
done

Then in VS Code, configure the MCP servers from mcp/mcp.linux.json (or mcp/mcp.wsl.json) into your VS Code MCP settings.

As a Copilot CLI plugin

copilot plugin install vandyG/power-bi-agent

MCP server setup

Add both servers to your VS Code or Copilot CLI MCP settings:

Linux / macOS (default):

{
  "servers": {
    "powerbi-modeling-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@microsoft/powerbi-modeling-mcp@latest", "--start"]
    },
    "microsoft-learn": {
      "type": "http",
      "url": "https://learn.microsoft.com/api/mcp"
    }
  }
}

WSL (so that Power BI Desktop on Windows is discoverable):

{
  "servers": {
    "powerbi-modeling-mcp": {
      "command": "cmd.exe",
      "args": ["/C", "C:\\Users\\<you>\\.vscode\\extensions\\analysis-services.powerbi-modeling-mcp-<ver>-win32-x64\\server\\powerbi-modeling-mcp.exe", "--start"]
    },
    "microsoft-learn": {
      "type": "http",
      "url": "https://learn.microsoft.com/api/mcp"
    }
  }
}

The Power BI Modeling MCP must run as a Windows process on WSL hosts; otherwise it cannot enumerate local Power BI Desktop instances.

Usage

In VS Code, open chat and select @power-bi (the consolidated agent). Topic skills auto-trigger from the user prompt; you can also invoke them by name:

@power-bi optimize this DAX measure for performance
@power-bi review my star schema for star-schema compliance
@power-bi which chart for showing churn cohort over time?
@power-bi convert this .sql to a Power Query expression

The agent always:

1. Connects to the live model (when relevant) before advising.
2. Checks for PBIP/PBIR/TMDL files when report or visual work can be done through source-controlled project files.
3. Searches Microsoft Learn before recommending non-trivial patterns.
4. Cites the documentation URLs it grounded the answer in.
5. Uses vscode_askQuestions if a Power BI Modeling MCP call returns a file/artifact URI, because the agent can only inspect it after the user saves it into the workspace.
6. Validates writes (dax_query_operations(operation: "Validate")) after creating measures.

PBIP workflow

When a workspace contains Power BI Project files, the agent can work directly against supported text metadata instead of treating report development as Desktop-only.

• Report-side editable surfaces: definition.pbir, PBIR definition/ folders, and already-registered files under StaticResources/RegisteredResources/.
• Report-side non-editable preview surfaces: report.json, mobileState.json, semanticModelDiagramLayout.json.
• Semantic model file workflow: prefer the Modeling MCP for live changes, but use *.SemanticModel/definition/ when the project is using TMDL and the task is explicitly file-based or source-control oriented.
• After external file edits, Power BI Desktop must be reopened or restarted before the changes appear.
• If a Modeling MCP result produces a file URI or vscode-chat-response-resource:// artifact, the user must save it to a workspace path before the agent can read it. If no file was generated because there was no data, the agent should continue without blocking.

Development

nix develop  # via the bundled shell.nix (or use direnv)

The repo currently has no automated tests; skills are validated by use. Update the version in plugin.json when shipping breaking changes to skill or agent definitions.

License

MIT — see LICENSE.

Initial agent/skill structure inspired by the awesome-copilot power-bi-development plugin and consolidated into a single chat-mode agent.