Power BI Agentic Development Marketplace
Teaches coding agents like Claude Code and GitHub Copilot Power BI and Microsoft Fabric, so you can automate semantic models, DAX, TMDL, and report work through skills, subagents, and hooks.
Hooks are individually toggleable via config.yaml, validation hooks claim to block only on structural errors, and the auto-upgrade hook is marked optional — basic least-privilege and confirmation design exists; but the etl skill advertises executing arbitrary PySpark and fabric-admin covers tenant-level remote operations, with no script-level evidence or confirmation mechanism shown for these high-power paths. Sensitive-data handling appears only as a bullet in the goblin-mode health-check description, no concrete implementation evidence.
The README is internally consistent: marketplace/plugin layout, Copilot CLI compatibility notes, hook version requirements (Copilot ≥1.0.26) and failure guidance ('Filename too long', hook troubleshooting commands) all cohere. A CI validation workflow exists but is gated behind vars.ENABLE_PLUGIN_CI and may not run by default — deducted.
Audience and scenario coverage is excellent (beginner onboarding through tenant governance), and environment fit is detailed (Windows long paths, dual .agent.md extension for Claude/Copilot, installation verification commands). Capability boundaries are partially stated (TE2 vs TE3, Pro/PPU suffices) but scattered — deducted one point.
Information architecture, install notes (two documented Copilot install paths plus an explicit warning that bare installs fail), known limitations (MAX_PATH, breaking transition, pin 26.25), a full GPL-3.0 LICENSE, version badge, and deprecation notes ('absorbs the old /audit-context') are all strong. Naming stability is explicitly disclaimed as weekly renaming/restructuring — deducted. Maintenance responsibility is attribution only (Kurt Buhler / Data Goblins) with no governance or response commitments.
The 'don't install every plugin' guidance directly improves usability; the te/pbir tool cascade and video walkthrough give concrete output paths. Actual output quality (DAX, reports) cannot be verified statically and cost/benefit claims are qualitative — deducted.
External dependencies are linked (pbir CLI, pbi-search, Microsoft docs, upstream issue tracker) and facts are mostly separated from inference. However, most skill behaviors are described without script or test content available for cross-checking; corroboration relies on external links only — deducted.
- The etl skill executes arbitrary PySpark/Python with full Delta table access, and governance skills perform remote tenant-level writes; review their scripts in an isolated environment before deployment.
- The author explicitly declares 26.26–26.38 a breaking transition; skills may be renamed or removed. Pin 26.25 or earlier for production.
- Windows users must enable long-path support at both OS and git level or installation fails.
- CI validation is gated behind the ENABLE_PLUGIN_CI variable; do not assume every change is plugin-validated.
- Publisher is registry-unverified and the README self-discloses AI-generated content; manually spot-check key skill files before adoption.
What does this agent do, and when should you use it?
This is an Anthropic-format plugin marketplace maintained by data-goblin, not a single plugin: the root contains only marketplace., with real capability split across eleven child plugins. Each plugin packages skills, subagents, and hooks: skills teach domain knowledge and activate automatically by task context or via manual /skill-name invocation; subagents are autonomous subprocesses for multi-step work such as semantic model auditing and BPA expression debugging; hooks are deterministic post-tool-use validation scripts that check TMDL, PBIR, DAX references, and .rdl structure. You register the marketplace with claude plugin marketplace add in Claude Code, then install child plugins via /plugin; Copilot CLI reads the same .claude-plugin/marketplace. without modification. Coverage spans Power BI Desktop connectivity, PBIP/TMDL/PBIR authoring, report and theme building, DAX and Power Query, Fabric CLI remote operations, tenant governance audits, and lakehouse ETL with Spark, Livy, and DuckDB. Releases are weekly, and versions 26.26 through 26.38 form a deliberate breaking transition in which skills may be consolidated, renamed, or removed.
Once installed, the agent gains concrete capabilities: connect, query, and modify a model in Power BI Desktop via the connect-pbid skill and Desktop Bridge, and capture DAX queries from visuals in real time with the query-listener subagent; author and validate TMDL and PBIR files directly (the pbip plugin ships a pbip-validator subagent plus hooks for PBIR, TMDL, and report binding validation); build a complete report from scratch with the pbir CLI — model discovery, pages, theme, visuals, field binding, filtering, publishing; write and optimize DAX and M/Power Query expressions; automate Tabular Editor work with BPA rules, C# scripting, and the te/te2 CLIs; run any remote Power BI or Fabric operation with the Fabric CLI (fab) on Pro or PPU, and audit tenant settings and delegated overrides; execute PySpark via ephemeral Livy sessions or query lakehouse data with DuckDB; and build Deneb/Vega-Lite, Python (matplotlib/seaborn), R (ggplot2), SVG (via DAX measures), and .pbiviz custom visuals, each backed by a reviewer subagent. Hooks fire automatically after Write/Edit or Bash commands touching relevant files, and each check can be toggled in plugins/pbip/hooks/config.yaml and plugins/pbi-desktop/hooks/config.yaml.
- A Power BI developer wants Claude Code to author and validate TMDL/PBIR files directly, catching structural syntax and model-binding errors automatically after every save.
- A BI team needs to audit and standardize an existing semantic model's naming conventions, memory usage, and DAX quality using the semantic-model-auditor subagent and lineage-analysis skill.
- A report author wants an agent to generate a complete interactive report from scratch following a design canon (visual hierarchy, color discipline, accessibility) and publish it via the pbir CLI.
- A Fabric administrator needs to audit tenant settings, delegated overrides, and Entra security group membership, or migrate workspaces from trial to production capacity, using fabric-cli and the audit-tenant-settings skill.
- A data engineer wants to run PySpark on Fabric Spark compute without creating a notebook artifact, or query lakehouse data with DuckDB for freshness and quality checks.
- A newcomer to agentic development wants a slow, jargon-free guided setup via the help-me-get-started skill in the goblin-mode plugin.
What are this agent's strengths and limitations?
- Unusually complete coverage: one marketplace spans the whole Power BI/Fabric lifecycle, from Tabular Editor BPA rules, C# scripting, and file-level TMDL/PBIR authoring to Fabric CLI tenant governance and Spark/DuckDB ETL.
- Hooks are deterministic validation, not LLM judgment: TMDL syntax, PBIR structure, DAX references, and .rdl files are verified after every write, each check individually toggleable via config.yaml, catching errors early.
- Dual-tool compatibility: skills load from skills/<name>/SKILL.md and agents use the dual *.agent.md extension, so the same repo works unmodified in both Claude Code and Copilot CLI.
- Purpose-built reviewer subagents add real depth: deneb-reviewer checks Vega/Vega-Lite syntax, svg-reviewer reviews SVG DAX measures, and query-listener captures DAX queries from Desktop visuals in real time.
- Weekly release cadence means frequent renaming and restructuring; versions 26.26–26.38 are a deliberate breaking transition, so you must pin 26.25 or earlier for stable behavior and cannot assume compatibility across transition releases.
- Strong coupling to the Power BI/Microsoft Fabric ecosystem and multiple external tools — pbir CLI, Tabular Editor, Fabric CLI (fab) — limits usefulness outside that stack.
- Copilot CLI hooks depend on the CLAUDE_PLUGIN_ROOT variable, set only in version 1.0.26 and later; older builds resolve hook commands to broken paths. Native Windows bash users may also hit a separate path-format bug (claude-code#11984), and Windows TMDL long paths require extra system-level configuration.
- Code is AI-generated with Claude Code, reviewed but potentially erroneous — the README says use at your own risk. The GPL-3.0 license forbids copying the skills into your own products, trainings, courses, or tools (including agent-rewritten copies) without attribution and a link to the original project.
How do you install or deploy this agent?
Prerequisites: Claude Code (or Copilot CLI; on Windows, TMDL repository paths can exceed 260 characters, so enable long paths at OS and git level, e.g. elevated git config --system core.longpaths true). Register the marketplace: claude plugin marketplace add data-goblin/power-bi-agentic-development. Then install plugins via /plugin in Claude Code, or from the command line, e.g.: claude plugin install goblin-mode@power-bi-agentic-development, claude plugin install tabular-editor@power-bi-agentic-development, claude plugin install pbi-desktop@power-bi-agentic-development, claude plugin install pbip@power-bi-agentic-development, claude plugin install semantic-models@power-bi-agentic-development, claude plugin install reports@power-bi-agentic-development. Copilot CLI users can run copilot plugin marketplace add data-goblin/power-bi-agentic-development followed by copilot plugin install tabular-editor@power-bi-agentic-development, or install a single plugin directly: copilot plugin install data-goblin/power-bi-agentic-development:plugins/pbip. Note: the bare copilot plugin install data-goblin/power-bi-agentic-development (no qualifier) installs nothing useful, because the root is a marketplace catalog, not a plugin.
How do you use this agent?
Skills activate automatically based on task context or can be invoked manually as /skill-name (e.g. /suggest-rule to generate BPA rules). In Copilot CLI, verify with /env, /plugin list, /skills list, /skills info pbip, and /agent. The pbi-desktop plugin requires a running Power BI Desktop instance to connect to; the reports plugin depends on the pbir CLI for report manipulation including live refresh and page screenshots; the fabric-cli plugin requires configured Fabric CLI (fab) credentials. Hooks run automatically after tool use; individual checks can be disabled by setting them to false in plugins/pbip/hooks/config.yaml (PBIR, TMDL, report binding validation) or plugins/pbi-desktop/hooks/config.yaml (DAX references, measure metadata, referential integrity, metadata cache). The README explicitly warns not to install every plugin: each skill competes for the agent's context window, so install per project need, remove when unneeded, and prefer project-scoped over user-scoped installs.