Power BI MCP Server: What It Is & How To Set It Up In 2026
A Power BI MCP server lets an AI client such as GitHub Copilot, VS Code or Claude read your Power BI semantic models and run DAX against them. As of October 2026 Microsoft offers three: Fabric IQ MCP (generally available, read-only, and the one Microsoft now recommends for asking questions), the older Power BI Consumption MCP server (preview), and the Power BI Authoring MCP server (formerly the Modeling MCP server), which changes models and runs either on your machine or as a hosted preview. All three act as the signed-in user (the local Authoring server can also use a service principal), so the real security question is what the AI does with your access and where the results go.
The Power BI MCP servers at a glance (October 2026)
Server | Status and where it runs | Sign-in and access | Can it change the model? |
|---|---|---|---|
Fabric IQ MCP: find reports and models, read schema, search values, run DAX | Generally available; hosted by Microsoft | Entra ID, delegated only ( | No, read-only |
Power BI Consumption MCP: read schema and report metadata, run DAX, generate DAX with Copilot | Preview; hosted by Microsoft | Entra ID; a Power BI admin enables a tenant setting; Build permission on the model | No write tools |
Power BI Authoring MCP (local): create and change tables, measures, relationships, RLS roles and translations | Version 1.0.0 on npm since September 25, 2026; runs on your machine (stdio) | Interactive sign-in or service principal; Write permission on the model | Yes, by default |
Power BI Authoring MCP (hosted): the same modeling work for models in Fabric workspaces | Preview; hosted by Microsoft | Entra ID as the signed-in user; the preview tenant setting must be on | Yes |
Sources: Microsoft's MCP servers overview, the Fabric IQ MCP guide, the Authoring MCP repository and its npm package. Names and status change often; check them before you roll anything out.
Which Power BI MCP server do you need?
- "What were sales by region last quarter?" Use Fabric IQ MCP. It finds the report or model by name (or from a browser URL you paste), reads the schema, lets your client's model write the DAX, runs it and returns the rows.
- "Add a year-over-year measure and document every table." Use the Authoring MCP server: the local one if the model is open in Power BI Desktop or saved as a PBIP/TMDL project, the hosted one if it lives in a Fabric workspace. Don't register both in the same client: their tools overlap, so the agent can't tell which to call. And don't use it to answer business questions; it runs DAX only so you can check the model you're building.
- An integration already built on the Consumption endpoint. Keep it running, but plan the move: the overview now names Fabric IQ as the server to prefer for semantic model consumption.
What each server can and cannot do
Fabric IQ MCP (read-only, generally available)
Fabric IQ MCP exposes six tools: DiscoverArtifacts, ResolveFabricItem, GetReportMetadata, GetSemanticModelSchema, ValueSearch and ExecuteQuery. It has no answering tool of its own: your client's model writes the DAX and explains the numbers, so answer quality depends on that model and on how well your semantic model is named and described.
The limits in Microsoft's setup guide (updated October 21, 2026):
- Reports and semantic models only. No dashboards, paginated (RDL) reports, Power BI apps, Fabric ontologies or data agents.
- One semantic model per
ExecuteQuerycall; the agent can combine results from separate calls. ExecuteQueryreturns 250 rows unless the client setsmaxRows. Larger results come back as an embedded CSV, which not every client displays.- Results are only as fresh as the model's last successful refresh.
- Delegated sign-in only, so it can't run as an unattended bot.
- Your tenant's home region must support all Fabric workloads; it isn't available in Power BI-only regions or sovereign clouds. The reports and models themselves don't need Fabric or Premium capacity, and you don't need a workspace role or Build permission: access to the report or model is enough.
Power BI Consumption MCP server (preview)
The first hosted server, at https://api.fabric.microsoft.com/v1/mcp/powerbi, offers Execute Query, Get Semantic Model Schema, Get Report Metadata and Generate Query. Per Microsoft's tool reference, schema and query calls need at least Build permission on the model, and Generate Query writes DAX with Copilot in Power BI, which needs a Copilot license for the user or organization; you can switch that tool off and let your client's model write the DAX. A Power BI admin has to enable the tenant setting for the MCP endpoint first (setup guide).
Power BI Authoring MCP server (local and hosted)
This is the server that writes. The local version is open source (MIT, plus a Microsoft EULA). It connects to a model open in Power BI Desktop, a model in a Fabric workspace, or a PBIP/TMDL folder, and exposes 21 groups of operations, from tables, measures and relationships to security roles, translations, DAX queries and traces. It can't touch report pages, visuals or diagram layouts.
The hosted version (preview) does the same modeling work on models in Fabric workspaces, with nothing to install, at https://api.fabric.microsoft.com/. It can't reach Power BI Desktop or files on disk, and transactions and traces stay local-only.
From the README: you need Write permission on the model; --readwrite is the default and asks for confirmation once per database before it applies an edit, while --readonly blocks every write; and the npm package blocks all tools until you accept the EULA, through an accept_eula call you authorize or with --accepteula.
How to set up Fabric IQ MCP (read-only questions)
Microsoft documents the setup with GitHub Copilot CLI, which comes with a preregistered Entra app:
- Install GitHub Copilot CLI and sign in. Microsoft also recommends the optional Fabric IQ skill from Skills for Fabric.
- Open
~/.copilot/mcp-config.json(on Windows,%USERPROFILE%\.copilot\mcp-config.json) and add the server:
{
"mcpServers": {
"FabricIQ": {
"type": "http",
"url": "https://fabriciq.svc.cloud.microsoft/v1/mcp/fabriciq",
"tools": ["*"]
}
}
}
- Start Copilot CLI and run
/mcp show FabricIQ. Sign in with your work or school account when the browser opens, and accept the requested permissions. - Test it: "Using the Sales Analytics report, show total sales by region for the latest quarter."
Two details that save time: don't put an Authorization header in the config, because a fixed token overrides the interactive sign-in; and if you build against the server, pin the tool contract by sending X-Variants: Fabric.Routing.FabricIQ.V1 on every request.
Other clients (VS Code, Claude, your own agent). Any client that supports Streamable HTTP and OAuth 2.0 can use the same endpoint. Without a preregistered Microsoft app, create a single-tenant Entra app registration with the delegated Power BI Service permissions Item.Read.All and Item.Execute.All; neither needs admin consent by default. With private links, the endpoint is https://api.fabric.microsoft.com/v1/mcp/fabriciq instead.
How to set up the Power BI Authoring MCP server
In VS Code: install GitHub Copilot Chat and the Power BI Authoring MCP extension, then check that powerbi-modeling-mcp is in Copilot Chat's tool list. On GitHub enterprise accounts, an admin first has to turn on "MCP servers in Copilot", which is off by default.
In Claude Desktop or another client that runs local servers: add the npm package (Node.js required). I'd start read-only and only drop the flag once you trust the workflow:
{
"mcpServers": {
"powerbi-authoring-local": {
"command": "npx",
"args": ["-y", "@microsoft/powerbi-modeling-mcp@latest", "--start", "--readonly"]
}
}
}
Then connect in plain language: Connect to '[File Name]' in Power BI Desktop, Connect to semantic model '[Semantic Model Name]' in Fabric Workspace '[Workspace Name]', or Open semantic model from PBIP folder '[Path to the definition/TMDL folder]'. For unattended runs, authenticate with a service principal (--authmode=serviceprincipal plus the AZURE_CLIENT_ID, AZURE_TENANT_ID and secret or certificate variables).
Security: what to check before you connect
These are the points I'd go through with whoever owns your Power BI tenant:
- Your data goes to the model provider. In the Authoring README's words: "AI assistance does not expand data access, but it may transmit accessed data to the LLM provider." Results from Fabric IQ take the same path, so check your client's model provider and its retention terms.
- Write access is the default on the Authoring server. One confirmation covers every later edit to that model, and Microsoft warns that "autonomous or misconfigured clients may perform destructive actions." Work on a copy, or on a PBIP folder under Git where a bad change is a diff you can revert.
- Admins can't switch off the local Authoring server on its own. It connects through the XMLA endpoint, and no tenant setting targets it; disabling XMLA is the only switch, and it cuts off every other XMLA tool too. Least-privilege workspace roles are the control that matters.
- Shared chats can leak. The README warns the model "might unintentionally expose sensitive information from the semantic model" in logs or responses.
- Tokens follow the allowlist. To reach Azure Analysis Services or SQL Server Analysis Services you add hosts to
PBI_MODELING_MCP_, and your access tokens are sent to those hosts. Only add endpoints your organization runs.ALLOWED_CONNECTION_HOSTS - Row-level security holds, per person. Fabric IQ runs every call as the signed-in user, with RLS and OLS applied and no administrator-impersonation path. The flip side: there is no shared service identity for a team bot.
- Data can carry instructions. Text stored in a model, such as a product description, is something your client's model reads. With a read-only server the worst case is a misleading answer; with the Authoring server in read-write mode, it could be an edit nobody asked for. Keep tool-call approval on.
The same split between "the server respects permissions" and "the client still sees everything you can see" applies to databases: see our Postgres MCP server guide, and ClickHouse's AgentHouse for a warehouse vendor's own MCP stack.
When GetDot is a better fit than a Power BI MCP server
A Power BI MCP server suits the person who already works in VS Code or Copilot and knows the model. It fits less well when the people asking are sales leads and finance managers in Slack or Teams: they don't run an MCP client, and nobody reviews the DAX their client's model wrote before the number lands in a deck.
That's the gap GetDot covers. Dot reads the same semantic model and runs DAX through Power BI, so answers use your measures (how Dot uses your Power BI semantic models). What it adds:
- Per-user rules on every query. Dot's row-level rules for a Power BI model (Team plan) are written in DAX and keyed on the asker's email or group. If row-level security is on for a model with no rule configured, Dot refuses the query instead of answering unfiltered.
- Answers where people already ask, in Slack, Teams, email or the web app, with the query shown so an analyst can check it (conversational analytics).
- Checked against your reports. Dot can rebuild a Power BI report as a Dot app, compare it with the original and report any remaining differences (October 2026 release notes).
- More than one model. Dot connects to 45+ data sources, so the same question can also reach the warehouse behind the report.
The trade-offs: Dot's connection runs as the Microsoft account that connected it, so RLS that lives in Power BI has to be mirrored in Dot's rules; Dot needs Premium, PPU or Fabric capacity, while Fabric IQ doesn't; and Dot never edits your model.
If your team likes MCP clients, you can have both: Dot is itself an MCP server at https://app.getdot.ai/ai/mcp for Claude, Cursor, GitHub Copilot in VS Code or Copilot Studio, and it answers with your own Dot permissions (Dot's MCP integration). See the Power BI integration and Power BI Copilot vs Dot, or start free with 300 one-time credits.
FAQ
Yes, three as of October 2026: Fabric IQ MCP (generally available, read-only), the Power BI Consumption MCP server (preview) and the Power BI Authoring MCP server, local or hosted. Community-built servers exist too.
Not for the servers themselves. Fabric IQ MCP needs an Entra work or school account and access to the report or model, not Fabric or Premium capacity; the Consumption server's Generate Query tool needs Copilot in Power BI; the Authoring server needs Write permission on the model.
Yes. Claude Desktop can run the local Authoring server through npx with the config above, and Fabric IQ MCP works with any client that supports Streamable HTTP and OAuth 2.0, possibly with a single-tenant Entra app registration.
Only the Authoring server can, and it writes by default. Start it with --readonly, keep the model backed up or in a PBIP project under version control, and keep tool-call approval on.
Fabric IQ MCP runs every call as the signed-in user, so Power BI row-level and object-level security restrict what comes back, and different users can get different answers to the same question.
Copilot works inside the Power BI interface; the MCP servers let other AI clients reach your models, with the client's own model writing the DAX. For what Copilot does and costs, see our Power BI Copilot guide.
Sven Rudolph
Sven is a co-founder of Dot. He writes about the technical side of AI analytics: MCP servers, BI integrations, and how Dot connects to the data stack.