How to Install Microsoft's Fabric Skills

0:05What a skill is

An AI coding tool such as Claude Code or GitHub Copilot CLI doesn’t come knowing Fabric’s APIs or query patterns. Microsoft’s skills for Fabric fill that gap. Each skill is a Markdown file, SKILL.md, that teaches the tool one Fabric job: the phrases that should trigger it, the REST APIs or commands to use, the authentication the job needs, and the practices to follow. Your tool loads the installed skills when a session starts, matches your prompt against them, and follows the instructions of the skill that fits. A skill doesn’t run anything on its own. The AI tool decides what to do with the guidance.

Microsoft Learn puts the difference between skills and MCP servers this way: skills teach the AI assistant what to do, and MCP servers do it. An MCP server is a running program that gives an agent live access to data and APIs, under the open Model Context Protocol.

	Skills	MCP servers
Purpose	Knowledge and patterns	Live access to data and APIs
Content	Markdown documentation	Executable servers
Runtime	Loaded into the AI’s context	Run as separate processes

The skills live in the microsoft/skills-for-fabric repository on GitHub, open source under the MIT license. Microsoft released them publicly at Build, on June 2–3, 2026, and Arun Ulag’s Build post calls them Agent Skills for Fabric. Claude Code and GitHub Copilot CLI install them from a plugin marketplace. VS Code picks them up from Copilot’s install, and Cursor, Windsurf, and tools that read AGENTS.md (Codex among them) use a clone of the repository. Claude Code is an officially supported target, and it’s the tool this article uses. If you haven’t used it with Power BI before, How To Use Claude Code for Power BI Development starts from the beginning.

Because a skill is plain Markdown, you can read exactly what it tells your AI before you install it. Here is the top of the fabriciq skill:

---
name: fabriciq
description: "Answers natural-language business questions over existing Power BI reports and semantic models through the FabricIQ MCP endpoint, discovering artifacts, resolving entities, generating the DAX itself, and returning a plain-language answer. Use when the user asks a data question, including one that says to query the model directly. For DAX the user supplies, or measures to save, use semantic-model-authoring."
---

The description, with the requests it names, is what your tool matches your prompt against. The rest of the file, and the reference files in the skill’s folder, carry the detail. A few skills also ship small helper scripts among their references.

2:29How one prompt reaches your workspace

This article follows one prompt from start to finish:

Please use the Fabric skills to document the Samples workspace.

On the way to an answer, it passes through every layer of the repository.

The layers between a prompt and your workspace, with the path this article’s example took.

Bundles are what you install. Each one packages skills with the MCP servers they use, and fabric-skills adds agents. There are two today, fabric-skills and powerbi-authoring, covered in the next section.

Agents are what Microsoft Learn calls experimental agent specializations: teammates with a persona and a set of workflows, which hand the detailed work to skills. The fabric-skills bundle installs four of them. FabricAdmin handles administration, governance, security, and workspace documentation and inventory, and its file describes it as “a pragmatic, security-conscious platform administrator”. FabricDataEngineer runs data engineering work across Spark, warehouses, pipelines, and lakehouses. FabricAppDev builds applications on Fabric data through Python, ODBC, XMLA, and REST. FabricIQ answers questions about Power BI reports and semantic models. You can name an agent in your prompt or let the tool choose one. Microsoft Learn’s example names it: Using FabricAdmin, document my Workspace FabricCLIDemo. The prompt above names no agent, and Claude Code chose FabricAdmin, whose file lists workspace documentation among its jobs and carries a workflow for it:

### Workspace documentation
Use the spark-cli skill to identify the workspace and use the other Fabric skills to operate.
When asked to "document my workspace" or similar, first make sure to confirm the workspace (find it and display its properties, Id, description).
Then, take a look at the Fabric workspace. 
Please document the data solution, the role of each artifact, the lineage, and what happens in which artifact with my data. 
Look at notebooks and pipelines to understand what they do.
BE CONCISE AND INTERESTING:
* Don't spell out all details (such as column names or types)
* See if there is any interesting business logic in views or stored procedures in warehouses or SQL Endpoints. Call out anything interesting
* Look at the semantic models to see what they do. 
* Write all results to a WorkspaceReport folder, in markdown format, with a top level overview plus one report per artifact type.

Skills each cover one Fabric item. Since August 2026 they’re named {item}-cli, such as sqldw-cli, spark-cli, eventhouse-cli, and powerbi-report-cli, and each one carries its authoring, consumption, and operations work as modes inside it. The repository has 25 skill folders.

The last layer is the way in. The skills tell the agent to call the Fabric and Power BI REST APIs, mostly through the Azure CLI’s az rest, and the bundles add MCP servers. fabric-skills brings three remote ones: FabricIQ (read-only questions over Power BI reports and semantic models), the hosted Power BI Authoring MCP server (in preview), and a Fabric SQL endpoint server. powerbi-authoring brings the local Power BI Authoring MCP server, the one The Power BI Modeling MCP Server, Hands-On works with; Microsoft renamed it from the Power BI Modeling MCP server in September 2026. Of the 25 skills, 24 also tell the agent to label every call to api.fabric.microsoft.com with the skill’s name. From sqldw-cli:

Every `api.fabric.microsoft.com` call must carry
`x-ms-fabric-skill: sqldw-cli` (`az rest`: `--headers "x-ms-fabric-skill=sqldw-cli"`),
including every LRO poll, `fabric_lro` and retry.

4:18Install the two bundles

Before you start, you need:

• An AI coding tool. This article uses Claude Code.
• The Azure CLI. Most Fabric operations sign in through it.
• Permission on the Fabric workspace you’ll point the skills at, for the operations you ask for.
• For powerbi-authoring, a Power BI tenant where you can create and publish semantic models and reports. Microsoft also recommends Power BI Desktop, to open and check projects locally.

The repository ships two bundles from one marketplace, named fabric-collection:

Bundle	Skills	Agents	MCP servers
fabric-skills	24, for authoring, consumption, operations, migration, and end-to-end architecture	FabricAdmin, FabricDataEngineer, FabricAppDev, FabricIQ	FabricIQ, the hosted Power BI Authoring MCP server, and the Fabric SQL endpoint server, all remote
powerbi-authoring	2: semantic-model-authoring and powerbi-report-cli	None	The local Power BI Authoring MCP server, started through npx

fabric-skills doesn’t include powerbi-authoring, so install both for full coverage. In Claude Code, type these three commands at the prompt:

/plugin marketplace add microsoft/skills-for-fabric
/plugin install fabric-skills@fabric-collection
/plugin install powerbi-authoring@fabric-collection

The first adds Microsoft’s marketplace; the other two install the bundles. Then quit Claude Code and start a new session, so it loads the skills. The same commands work in GitHub Copilot CLI. From a terminal instead of a Claude Code session, the repository’s MCP guide uses claude plugin marketplace add and claude plugin install with the same arguments.

Plugin bundles install whole. To install a single skill, Microsoft points to APM, the Agent Package Manager:

apm install microsoft/skills-for-fabric --skill sqldw-cli --target claude

The --skill values are the folder names under skills/ in the repository. Two things come along with any single skill: APM registers every Fabric MCP server unless you add --only apm, and it deploys all the repository’s agents. A smaller install keeps the AI focused and limits what it can reach, which matters most when it runs unattended, in a release pipeline for example.

The bundles have changed since the recording in June 2026. The three focused bundles fabric-authoring, fabric-consumption, and fabric-operations were retired in August 2026, and their names now install the full fabric-skills bundle, so an existing install keeps working. The --filter option for narrowing the full bundle left the README in July 2026, and single-skill installs through APM arrived in September. Microsoft Learn’s install page, last updated September 29, 2026, still lists the retired bundles and --filter; the repository’s README has the current commands.

Sign in once, then check the connections

The skills and the bundled MCP servers sign in through the Azure CLI. Check which tenant it’s on, without printing any credentials:

az account show --query tenantId --output tsv

If you aren’t signed in, or it’s the wrong tenant, sign in:

az login --tenant <tenant-id> --allow-no-subscriptions

Since September 2026, the fabric-skills plugin reuses that sign-in for its three remote MCP servers in local Claude Code, so there’s no token to copy and no app to register. Your workspace permissions, tenant policies, and consent settings still apply.

Then check the servers. From a terminal, run claude mcp list; inside Claude Code, type /mcp. Read each server’s status: the repository’s MCP guide warns that a successful install, or an exit code of zero, doesn’t mean every MCP server connected.

If you registered a FabricIQ server by hand, that older entry can take precedence over the plugin’s, and FabricIQ moved to a new endpoint on September 17, 2026. If the old entry fails, remove only that entry, using the name and scope /mcp shows, for example:

claude mcp remove --scope local FabricIQ

Document a workspace with one prompt

1:02 With the bundles installed and the Azure CLI signed in, the video runs the article’s example prompt against a workspace named Samples, which holds Power BI’s eight standard sample domains: a semantic model, a report, and a dashboard for each.

One prompt. Claude Code picked the FabricAdmin agent, which took 14 tool calls and just under four minutes.

The agent finished in 14 tool calls, 56.6k tokens, and 3 minutes 51 seconds. It summarized the workspace in the conversation (24 items, 8 of each type, every report bound to the model of the same name) and wrote the full documentation to C:\Demo\Samples-workspace-documentation.md.

The top of the file the agent wrote.

The file opens with an overview of the workspace and its capacity, then a table of which semantic model each report is bound to, then an inventory by item type, with the main tables of each model read live through DAX’s INFO.VIEW.TABLES(). An appendix at the end of the file records how each part was collected:

• Workspace and item inventory: the Fabric REST API, GET /v1/workspaces and /v1/workspaces/{id}/items
• Report-to-model bindings: the Power BI REST API, GET /v1.0/myorg/groups/{id}/reports
• Model configuration and refresh status: the Power BI REST API, .../datasets and .../datasets/{id}/refreshes
• Table structure: DAX EVALUATE INFO.VIEW.TABLES() through /datasets/{id}/executeQueries

You could have the agent regenerate this documentation on every release to production, or every night, and see what meaningfully changed from one run to the next.

Your own run may look different. The FabricAdmin workflow asks for a WorkspaceReport folder, with an overview plus one report per item type, and this run wrote a single file instead. Microsoft Learn’s version of the prompt is longer and spells out what to look at, which suits a workspace with notebooks, pipelines, and warehouses in it:

Document my Fabric Workspace: take a look at the workspace called
FabricCLIDemo. Document the data solution, the role of each
item, the lineage, and what happens in each item with my
data. Look at notebooks, pipelines, views, stored procedures, and
semantic models. Write all results to a WorkspaceReport folder in
Markdown format.

How much to let the agent do without asking

The demo started Claude Code with --permission-mode bypassPermissions, so it ran the whole job without stopping to ask before an API call or a file write. Claude Code’s documentation says to use that mode only in isolated environments, like containers or VMs, where Claude Code can’t cause damage. The demo ran in a Windows virtual machine in Parallels, and this prompt only reads from Fabric. Other skills in the repository create and change items, and an MCP client acts with your Fabric permissions; Microsoft’s Power BI MCP overview warns that “an autonomous or misconfigured client might perform destructive actions.”

For a first run, Microsoft Learn recommends planning before executing: start in plan mode, check the plan, and only then let the agent act. In Claude Code’s plan mode, Claude reads and explores but doesn’t edit your files. Learn also suggests naming the workspace in every prompt and adding “Before making any changes, ask me which workspace to use.”

My own view is that this comes down to trust. We hire people and trust them every day to avoid serious mistakes and to act in our best interest. Yet we tend not to extend that trust to AI, even though it can be tested, constrained, and given guardrails, and it has no motive to cheat or steal. Things can still go wrong, and some third parties will try to exploit an agent, through prompt injection for example, so you need to protect against that. In the end, your company will likely set the policy, and most people won’t be able to turn on bypass mode.

Claude Code has a setting for that policy: an administrator who sets disableBypassPermissionsMode in managed settings turns bypass mode off, and no user setting can override it.

5:33Keep up with a fast-moving repository

The repository was on version 0.3.3 when the video was recorded in June 2026, with about 520 stars. By October 17, 2026, the latest GitHub release was v0.3.18, from September 25, the changelog already listed 0.3.19, and the repository had 1,252 stars. Releases have come about once a week.

Turn on automatic updates in Claude Code once. Microsoft’s marketplace is a third-party marketplace, so it doesn’t update on its own:

1. Type /plugin.
2. Open Marketplaces and select fabric-collection.
3. Choose Enable auto-update.

To update on demand from a terminal, run claude plugin update fabric-skills@fabric-collection, and the same for powerbi-authoring. Administrators can set "autoUpdate": true on the marketplace’s entry in managed settings instead.

Read CHANGELOG.md before you rely on a skill’s name in a prompt or a script. Names change: on September 17, 2026, powerbi-report-cli replaced four separate report skills.

6:06What else the skills cover

Beyond documentation, each of these jobs has a skill behind it:

• Building Power BI reports and semantic models: powerbi-report-cli and semantic-model-authoring. How To Use Microsoft’s Fabric Skills To Build a Power BI Report shows the Power BI skills building a report. AI Can Now Check Its Edits in Power BI covers the Desktop Bridge, which lets an agent reload the report in Desktop and screenshot it.
• Asking plain-English questions of your Power BI data: fabriciq.
• Diagnosing slow queries and warehouse health: the operations mode of sqldw-cli.
• Real-time intelligence: eventhouse-cli, eventstream-cli, and activator-cli.
• A medallion architecture, built end to end: e2e-medallion-architecture.

Install both bundles, sign in with az login, and run the workspace prompt against a workspace you can read. Then open the file it writes and check it against what you know is in that workspace.