Microsoft's powerbi-authoring plugin bundles agent skills and an MCP server so coding agents can build Power BI semantic models and PBIR reports with guardrails. A GitHub trending entry pointed at microsoft/powerbi-authoring-plugin, which 404s; the supported path is the powerbi-authoring plugin inside skills-for-fabric.
Microsoft Learn's agentic overview recommends this plugin as the fastest on-ramp: one marketplace install wires modeling MCP plus authoring skills instead of manual server registration.
What installs in one command
From Learn and the v0.3.3 release notes:
| Component | Role |
|---|---|
semantic-model-authoring | Import, DirectQuery, Direct Lake models; DAX; deploy/refresh |
powerbi-report-planning | Requirements → locked spec before PBIR work |
powerbi-report-design | Visual design contract before files |
powerbi-report-authoring | PBIR/PBIP edits via powerbi-report-author CLI |
powerbi-report-management | Fabric REST report lifecycle via az rest |
| Power BI Modeling MCP | @microsoft/powerbi-modeling-mcp preconfigured |
Microsoft's MCP authoring doc stresses tools without skills are insufficient—agents need the how (naming, sequencing, validation) bundled in the plugin.
Dashboard metrics tiles reflecting Power BI agent authoring workflows
Install for GitHub Copilot CLI
Learn documents two CLI steps: copilot plugin marketplace add microsoft/skills-for-fabric, then copilot plugin install powerbi-authoring@fabric-collection.
Then run /skills and confirm semantic-model-authoring appears.
Agent skills library motif for Fabric marketplace plugin bundles
Analogous installs exist for other hosts that support the Fabric marketplace; our FoundryX plugin guide shows how Microsoft packages skills + MCP for different agent surfaces.
Skill workflow: planning before pixels
Microsoft's v0.3.3 release notes emphasize powerbi-report-planning as a gate: agents must produce a locked spec with explicit human approval before powerbi-report-authoring touches PBIR files. That pattern mirrors legal Copilot skills at Ignite—high stakes domains get documented checkpoints.
The design skill ships 19 references (accessibility, page archetypes, typography). Agents are instructed to call powerbi-report-design when visuals are ambiguous, reducing dashboard soup.
Semantic models vs reports vs FabricIQ
| Skill | When Microsoft docs say to use it |
|---|---|
semantic-model-authoring | Tables, DAX, relationships, deployment |
powerbi-report-authoring | PBIR layout after spec/design |
fabriciq | Natural-language Q&A on published artifacts |
semantic-model-consumption | Raw DAX you already wrote |
Mixing them wrong—asking FabricIQ to rename measures—violates skill boundaries and burns tokens.
How this fits Ignite 2026
Ignite emphasized Cowork for finance with Excel + ERP skills and Dynamics 365 in Copilot. The Power BI plugin is the builder-facing mirror: agents that author datasets and PBIR before humans publish to Fabric capacities.
Pair with:
- Copilot Managed Runtime for custom analytics apps.
- Blender MCP addon as another MCP-vertical pattern.
Risks and governance
Permission boundary motif for scoped MCP access to Fabric workspaces
- Over-published DAX — agents can break refresh schedules or leak measures; use dev workspaces first.
- REST management skills —
powerbi-report-managementcan delete artifacts; scope service principals narrowly. - Credential paths — MCP servers inherit Desktop or Fabric tokens; rotate keys if agents run on shared runners.
What data teams should do this week
- Pilot on a PBIP repo with version control before letting agents touch shared datasets.
- Force planning skill for net-new dashboards so layout commits happen only after spec lock.
- Log MCP tool calls alongside Agent 365 policies if Copilot routes these skills in production.
PBIR, PBIP, and CLI dependencies
Report skills assume familiarity with Power BI Project layouts and the powerbi-report-author CLI referenced in skill docs. Agents should run powerbi-desktop validation steps Microsoft lists before publishing—otherwise render-only bugs slip to executives.
For Direct Lake models, semantic-model skills include Direct Lake guidelines references in the changelog; verify Fabric capacity SKU supports Direct Lake before agents migrate import models.
Manual path without the plugin
Microsoft Learn still documents registering MCP servers manually and importing individual skills from the Fabric collection. The plugin exists because onboarding friction killed early agent pilots—one command registers @microsoft/powerbi-modeling-mcp and loads semantic-model-authoring references.
Teams allergic to GitHub Copilot CLI can port skills to other hosts that read the same SKILL.md files, but must replicate MCP registration themselves.
Reference library size and maintenance
The v0.3.3 changelog lists dozens of reference markdown files for report design and authoring. Agents with small context windows may need skill pruning—invoke design or authoring skills individually rather than loading the entire plugin when only DAX edits are required.
Watch check-updates skill bundled in the plugin; Microsoft may ship breaking PBIR schema guidance without a headline blog post.
Corporate data governance checklist
- Label sensitivity on semantic models before agents connect.
- Restrict service principals to dev workspaces.
- Require four-eyes approval on production dataset refresh schedules agents modify.
- Archive agent prompts that contain DAX or business logic under your model risk policy.
skills-for-fabric release tags accurate as of November 18, 2026.
Example agent prompts that fit Microsoft's skills
These prompts align with documented skill boundaries—adapt to your tenant:
- "Using semantic-model-authoring, add a DAX measure for rolling 28-day active users in the dev workspace semantic model, then document the definition in the model glossary."
- "Run powerbi-report-planning only: produce a one-page spec for an executive dashboard from the Sales semantic model; stop before writing PBIR."
- "After I approve the spec, invoke powerbi-report-authoring to implement page one with accessible colors per powerbi-report-design references."
Bad prompts that violate skill docs:
- "Email the CFO the dashboard" — outside Power BI authoring scope.
- "Delete production datasets" — should fail governance checks.
Fabric capacity and performance notes
Agents editing DirectQuery models can trigger source system load. Throttle agent runs during month-end close windows when finance teams already stress ERP connectors. Microsoft's Ignite finance Copilot story and this plugin intersect where Excel + semantic models feed executive narratives.
Telemetry and cost control for agent authors
Each MCP tool call may hit Power BI Desktop, Fabric REST, or Azure CLI paths. Log:
- Workspace ID targeted
- Skill name invoked
- Human approval timestamps on planning outputs
- Token spend on the outer coding agent hosting the plugin
Without telemetry, finance sees a Copilot bill without knowing which dashboard agent caused DAX refresh storms.
Partner and SI delivery angle
Systems integrators packaging Fabric + GitHub Copilot should ship runbooks that name this plugin explicitly—clients confuse generic "Copilot" with Power BI-specific skills. Include rollback steps when agents publish invalid PBIR (Power BI Report) JSON.
Licensing and support boundaries
Microsoft's open skills-for-fabric repo is source-available skills, not a paid SKU. Support tickets still flow through Fabric or Power BI channels depending on failure mode. Document internal SLAs before agents touch production executive dashboards.
Comparison to other Microsoft agent plugins
| Plugin family | Primary artifact | This plugin |
|---|---|---|
| FoundryX | Azure AI Foundry agents | Different surface |
| Copilot Managed Runtime | Custom web apps | Complementary |
| powerbi-authoring | Datasets + PBIR | This post |
Choose one orchestrator (Copilot, Claude Code, or custom) to avoid duplicate MCP registrations fighting for Power BI Desktop handles.
Testing matrix before production
| Test | Pass criteria |
|---|---|
| Dev workspace publish | Report renders in Desktop |
| DAX measure add | Refresh succeeds |
| Invalid relationship | Agent catches error |
| Production block | Agent cannot bypass RBAC |
Run the matrix weekly while model weights underneath the coding agent change.
Why GitHub trending mislinked the repo
Feed aggregators sometimes invent powerbi-authoring-plugin slugs that do not exist as standalone repositories. Always verify microsoft/skills-for-fabric paths before wiring CI bots to clone the wrong URL. The 404 on the trending link is a discovery hazard, not proof Microsoft withdrew the product.
Upskilling path for analysts
Analysts who historically clicked in Desktop can pair this plugin with explainx.ai workshop-style learning on Copilot + Fabric literacy before they delegate report builds to agents. Human reviewers should still sign visual encodings—agents excel at first drafts, not brand compliance.
Version skew between Desktop, Fabric, and skills
Power BI monthly updates can desync skill references from your tenant features. Pin skills-for-fabric release tags in internal docs and rerun smoke tests after Microsoft monthly drops—especially when TMDL or PBIR schema changes hit.
Observability hooks
Export MCP audit logs to the same SIEM stream as Windows agent telemetry if agents run on VDI with Power BI Desktop. Correlating agent identity with dataset refresh events catches runaway automation early.
File issues on skills-for-fabric when agent skills disagree with Desktop UI labels—upstream fixes help every Fabric customer, not only your tenant.
Schedule quarterly refresh training for analysts anytime Microsoft adds a new powerbi-report-* skill—agents will invoke them automatically once your marketplace pin updates.
Treat Fabric capacity pauses like production freezes: disable agent plugins before Black Friday or fiscal close when refresh storms could trip source database guardrails. Communicate freeze dates in the same channel where you announce Copilot Autopilot pilots so teams do not schedule conflicting automation.
Archive before and after PBIR in git when agents author reports so compliance can reconstruct who changed executive metrics.
Pair plugin adoption with Blender MCP lessons: both show Microsoft and partners shipping vertical MCP bundles instead of expecting users to wire tools manually every time.
When executives ask for single-pane agent governance, point them to Agent 365 dashboards first, then this plugin as the Power BI slice of a broader Fabric automation strategy—not the entire agent program by itself or a replacement for trained human analysts and reviewers on every deliverable.
