Docs  \  AI  \  AI tools  \ 

Set up Codex CLI for agentic development with Power BI Desktop

This article describes how to configure Codex CLI so that an AI agent can read and modify a Power BI semantic model open in Power BI Desktop, and the report of a Power BI project.

The only requirement we assume is Power BI Desktop, already installed. Everything else is part of this setup.

Sample model used in all the steps: ContosoDemo10k.zip. Please, download it before you start.

At the end you have an agent in the terminal that reads and writes the semantic model, and that also creates and validates report pages.

One note on the name, because it changed around the product and not in it. The client is still called Codex CLI, and the package is still @openai/codex. What moved is the documentation, from developers.openai.com/codex to learn.chatgpt.com/docs, and the desktop application, which is now part of the ChatGPT desktop app. The command you type is codex.

Requirements

  • Power BI Desktop, installed.
  • Node.js 22 or later. The Power BI Modeling MCP server is started with npx, which is part of Node.js. Codex CLI itself does not need it, unless you install it with npm. A new machine does not have it, so step 1 installs it.
  • A ChatGPT account. Codex CLI signs in with it, and the plan you have determines the limits of the session. Check the plans page before you start.
  • Git for Windows. The repository of the skills is downloaded with git, and Windows does not include it. Step 11 installs it.
  • Write permission on any semantic model you modify. The MCP server follows the same rules as the Power BI external tools.

MCP stands for Model Context Protocol. The Power BI Modeling MCP server runs on your machine and connects to Power BI Desktop like an external tool. Codex CLI is the client that hosts the agent, and it starts the server as a local process.

Back up your model before an agent writes to it. With the sample model, extract the archive again if something goes wrong.

If your organization manages ChatGPT centrally, an administrator can restrict the MCP servers with an allow list, and can disable them completely. In that case the configuration below is accepted and no tool appears.

Step 1: install Node.js

A new installation of Windows does not have it. npm and npx are part of Node.js, and the Power BI Modeling MCP server is started with npx.

Open the command prompt

Several steps of this article ask you to type a command. Windows has two programs that run the commands, and they are not interchangeable:

  • The command prompt. Press Windows+R, type cmd, and press Enter. The window shows a folder path followed by >.
  • PowerShell. Press Windows+R, type powershell, and press Enter. The line starts with PS.

Every command of this article says which of the two to use. To open a window already positioned in a folder, open that folder in File Explorer, type cmd in the address bar, and press Enter. When an installation asks for administrator rights, find Command Prompt in the Start menu, right-click it, and choose Run as administrator.

Install Node.js

Choose one of the two methods. LTS stands for Long Term Support.

Option 1: the installer. Download it from nodejs.org/en/download and run it.

Option 2: WinGet. In the command prompt:

winget install --id OpenJS.NodeJS.LTS --source winget

Both methods write in the folders of the machine, so they ask for administrator rights. If you do not have them, use one of the options at the end of this step.

Close the command prompt and open it again, so that it reads the updated path. Then verify the installation, in the command prompt:

node --version

The command returns the version you installed. The LTS installer provides 22 or later.

Install Node.js without administrator rights

These are options, not a replacement of the installer. Use one of them when the machine is managed and you do not have administrator rights. Each one writes in your user profile, and each one provides npm and npx.

Option 1: the standalone binary. The download page offers a standalone archive besides the installer. Extract it in a folder of your user profile, for example %LOCALAPPDATA%\nodejs, and add that folder to the Path variable of the user, in Settings > System > About > Advanced system settings > Environment variables. Nothing else is installed.

Option 2: fnm, the Fast Node Manager. It installs and switches the versions of Node.js in your user profile. Install fnm with WinGet, in the command prompt:

winget install Schniz.fnm

Add this line to your PowerShell profile, so that every session finds Node.js. The command notepad $PROFILE, typed in PowerShell, opens that file:

fnm env --use-on-cd --shell powershell | Out-String | Invoke-Expression

Open a new PowerShell window, install the LTS version, and read the version number it installed:

fnm install --lts

fnm list

Then set that version as the default, replacing <version> with the number of the LTS entry of the list:

fnm default <version>

Option 3: Scoop. It is a package manager that installs in %USERPROFILE%\scoop. Install it in PowerShell, in a window that is not elevated:

irm get.scoop.sh | iex

Then install Node.js, in the same PowerShell window:

scoop install nodejs-lts

With any of the three options, close the window, open a command prompt, and verify with node --version and npm --version.

Step 2: install Codex CLI

Choose one of the three methods.

Option 1: the installer of OpenAI. The command starts PowerShell itself, so it runs in the command prompt:

powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

Then close the command prompt window used for the installation and open a new one to open Codex CLI.

Option 2: npm. In the command prompt:

npm install -g @openai/codex

Option 3: WinGet. NOTE: as of September 2026, WinGet setup does not create the link to codex, so you should manually create it. Use any of the previous other options if possible. In the command prompt:

winget install -e --id OpenAI.Codex

There is no Microsoft Store version of Codex CLI. The Store contains the ChatGPT desktop app, which is a different product and includes its own copy of the agent.

Codex CLI runs natively on Windows, in command prompt or PowerShell. The Windows Subsystem for Linux is an option, not a requirement.

Verify the installation, in the command prompt:

codex --version

Step 3: open the sample model

  1. Download ContosoDemo10k.zip.
  2. Extract the archive in a local folder, for example C:\Demo. Please, do not open the file directly from the compressed folder.
  3. Open ContosoDemo10k.pbix in Power BI Desktop and leave Power BI Desktop open.

The title bar shows ContosoDemo10k. You use that name in step 6.

Step 4: register the Power BI Modeling MCP server

There is no extension to install. The server is an npm package, and Codex CLI starts it on demand.

Run this command in the command prompt:

codex mcp add powerbi-modeling-mcp -- npx -y @microsoft/powerbi-modeling-mcp@latest --start --readonly

The --readonly argument blocks every write operation. We suggest it for the first session, so that a wrong prompt cannot modify anything. Step 8 replaces it.

The command writes the entry in %USERPROFILE%\.codex\config.toml. You can edit that file directly instead of running the command:

[mcp_servers.powerbi-modeling-mcp]
command = "npx"
args = ["-y", "@microsoft/powerbi-modeling-mcp@latest", "--start", "--readonly"]
startup_timeout_sec = 60

The default timeout is 10 seconds, and the first start of npx downloads the package, so raise it as in the example.

Verify the registration, in the command prompt:

codex mcp list

Step 5: start the agent and sign in

  1. Open a command prompt in the folder that contains the sample, for example C:\Demo.

  2. Start the client:

    codex
  3. Answer the trust prompt for the folder. Until you trust it, Codex CLI stays in read-only mode and does not read the configuration of the project.
  4. Sign in, choosing Sign in with ChatGPT. The command codex login does the same from the terminal.

A browser window opens, you sign in, and the control returns to the terminal.

Step 6: connect to Power BI Desktop

Send this prompt:

Connect to 'ContosoDemo10k' in Power BI Desktop

The answer reports the model name and an active connection. The first call to the MCP server raises a confirmation prompt. Read it: it is the only checkpoint before a change.

Step 7: verify the connection

Send these two prompts to the agent, one after the other:

List the tables and their row counts
Show me the relationships in the model

Both answers arrive in a few seconds. The chain works: Codex CLI, MCP server, Power BI Desktop.

Step 8: enable the write operations

The session started in read-only mode, so the next prompt would fail. Remove the restriction:

  1. Exit the client.

  2. Remove the read-only registration, in the command prompt:

    codex mcp remove powerbi-modeling-mcp

  3. Register the server again without --readonly:

    codex mcp add powerbi-modeling-mcp -- npx -y @microsoft/powerbi-modeling-mcp@latest --start

  4. Start the client again, allowing the writes in the folder:

    codex --sandbox workspace-write --ask-for-approval on-request
  5. Connect again, as in step 6.

Then send this prompt:

Create a measure that returns the Sales Amount of the previous year using the same format of the original measure.

The measure appears in Power BI Desktop without a refresh. Each write operation raises a confirmation prompt. The same approach applies to bulk operations, like format strings and display folders on hundreds of objects in a single request.

The sandbox of Codex CLI has three levels, read-only, workspace-write, and danger-full-access, and the approval policy has three values, untrusted, on-request, and never. The pair used above is the one we suggest for this scenario. The older --full-auto argument is deprecated.

Step 9: enable the preview features for the report layer

The MCP server operates on the semantic model. The report layer requires a different tool, the Power BI report authoring skill, which works only on PBIP files in PBIR format. PBIP stands for Power BI Project, PBIR for Power BI enhanced report format. The skill cannot modify a .pbix file.

In Power BI Desktop, open File > Options and settings > Options > Preview features and enable these three options:

  • Power BI Project (.pbip) save option
  • Store reports using enhanced metadata format (PBIR)
  • Enable external tool access to Power BI Desktop through secure local APIs

The third option is the Power BI Desktop Bridge, and it is enabled by default. The report authoring skill uses it to reload the project and to capture a screenshot of a page, so that the agent verifies its own work. Verify that the option is checked.

Restart Power BI Desktop.

Step 10: save the sample as a project

Use File > Save as and choose the Power BI Project (*.pbip) file type. Power BI Desktop creates this structure:

ContosoDemo10k.pbip
ContosoDemo10k.SemanticModel/
ContosoDemo10k.Report/
.gitignore

The model is a set of TMDL files (Tabular Model Definition Language) and the report is a set of JSON files. Put the folder under source control and commit a baseline, so to undo a wrong operation with one command.

Step 11: install the report authoring skill

Git is required for this step. The repository is downloaded with git, and Windows does not include it. Install it, then close the command prompt and open it again.

Choose one of the two methods.

Option 1: the installer. Download it from git-scm.com/downloads/win and run it.

Option 2: WinGet. In the command prompt:

winget install --id Git.Git --source winget

The skills call two command-line tools. Install them in the command prompt:

npm install -g @microsoft/powerbi-report-authoring-cli@latest @microsoft/powerbi-desktop-bridge-cli@latest

The skill is part of the powerbi-authoring plugin of the skills-for-fabric repository, published by Microsoft. The plugin contains five skills: semantic model authoring, report planning, report design, report authoring, and report management.

Codex CLI is not one of the clients with a marketplace command for these skills. Microsoft documents the manual path for it: clone the repository, and the AGENTS.md file in the root is detected automatically. Codex CLI reads AGENTS.md from the folder of the project and from the folders above it.

Clone the repository next to the sample, in the command prompt:

git clone https://github.com/microsoft/skills-for-fabric.git C:\Demo\skills-for-fabric

Codex CLI also loads the skills it finds in %USERPROFILE%\.agents\skills, in the same SKILL.md format used by the repository. Copy the skills of the plugin there, in the command prompt:

xcopy /E /I "C:\Demo\skills-for-fabric\plugins\powerbi-authoring\skills" "%USERPROFILE%\.agents\skills"

Please, verify the folder names in the repository before you copy, because the layout changes between versions. The repository also has a skills folder in its root, which contains all the skills of the collection. The five skills of the report layer are the ones in the folder of the plugin. Restart the client, and ask the agent to list the skills it can use.

This combination is not documented by Microsoft or by OpenAI, and it can stop working with any update of the repository. Treat it as a manual installation and verify the result on the sample before you use it on a model that matters.

Step 12: connect the agent to the project

Start the client in the command prompt, in the folder that contains the .pbip file, so that the agent sees both the model folder and the report folder:

codex --cd "C:\Demo"

Then send this prompt:

Open semantic model from PBIP folder 'C:\Demo\ContosoDemo10k.SemanticModel'

The MCP server manages the semantic model folder. The authoring skill reads and writes the report folder as files.

Step 13: create a report page

Create a report page with a line chart showing Sales Amount by Quarter, and a card showing the Sales Amount of the last year that has data.

Then ask the agent to validate the report, which checks the structure of the PBIR files, and open the .pbip file in Power BI Desktop. The skill runs this command, and you can run it yourself in the command prompt:

powerbi-report-author validate "C:\Demo\ContosoDemo10k.Report"

Save any manual change in Power BI Desktop before the agent iterates. The agent reads the files on disk and does not see the unsaved state. Editing in both places at the same time loses one set of changes.

The examples in the documentation use cards, bar charts, clustered column charts, tables, KPI cards, and slicers, and the skill converts the legacy card and matrix visuals into the modern cardVisual and pivotTable. There is no published list of the supported visuals, so expect some trial and error. Q&A, Bing maps, and filled maps are announced for deprecation, and Microsoft recommends avoiding them.

Configuration options

The command-line arguments of the MCP server go after -- in the codex mcp add command, or in the args array of the TOML section.

Option Default Description
--start required Starts the server.
--readwrite enabled Allows write operations, each one with a confirmation.
--readonly   Blocks all the write operations.
--skipconfirmation   Removes the confirmation prompts.
--compatibility PowerBI Set it to Full for Analysis Services databases.
--authmode interactive Set it to serviceprincipal for unattended scenarios.

For the service principal authentication, set AZURE_CLIENT_ID and AZURE_TENANT_ID in the environment of the server, with either AZURE_CLIENT_SECRET or AZURE_CLIENT_CERTIFICATE_PATH.

On a model that matters, run the first session with --readonly, as in step 4.

Alternative installation methods

Manual installation. Download the VSIX package of the Visual Studio Code extension of the MCP server, rename it with the .zip extension, extract it, and configure the path of the executable. This avoids the download that npx performs at every start:

[mcp_servers.powerbi-modeling-mcp]
command = "C:\\MCPServers\\PowerBIModelingMCP\\extension\\server\\powerbi-modeling-mcp.exe"
args = ["--start"]

If npx does not start on Windows, use the command interpreter as the launcher:

[mcp_servers.powerbi-modeling-mcp]
command = "cmd"
args = ["/c", "npx", "-y", "@microsoft/powerbi-modeling-mcp@latest", "--start"]

The npx command is a .cmd script on Windows, and a process that starts it without a shell can fail to find it. Write this form directly in the file, because the codex mcp add command does not always preserve the /c argument.

Limitations

  • The MCP server operates on the semantic model, not on the report.
  • The report authoring skill operates on PBIP projects in PBIR format, not on a .pbix file.
  • Neither of them can do anything your permissions do not allow, because they operate with your identity.
  • Model metadata does not stay local: table names, column names, measure definitions, and query results are sent to the language model of the service you signed in to.
  • The MCP server, the report authoring skill, and the Power BI Desktop Bridge are all in preview. Behavior and tools can change before general availability.
  • The agent proposes and executes. The review is your responsibility.
  • The report authoring skill has no marketplace installation for Codex CLI. The manual path of step 11 is not covered by the documentation of Microsoft or of OpenAI.
  • The sandbox on Windows depends on the version of the client and on the policy of the machine. Verify which mode is active before you rely on it.

Troubleshooting

The codex command is not found. Close and open the terminal again, so that it reads the updated path. With npm, verify that the global folder of npm is in the path.

The MCP server does not appear in the list of tools. Run codex mcp list in the terminal. Then verify, in this order: the entry exists, you trusted the folder, your organization does not restrict the MCP servers.

The server times out at the start. Raise startup_timeout_sec in the TOML section. The default is 10 seconds, which is not enough for the first download of the package.

The server does not start and the error mentions npx. Use the cmd form of the alternative installation methods, or the absolute path of the executable.

The agent does not read the configuration of the project. Trust the folder. Until then, Codex CLI ignores the project configuration and stays in read-only mode.

The npm or the npx command is not found. Node.js is not installed, or the terminal was opened before the installation. Close the terminal, open it again, and run node --version. See step 1.

The git command is not found. Git is not installed, or the command prompt was opened before the installation. Install Git for Windows as in step 11, then close the command prompt and open it again.

The server starts and provides no tools. Verify that Node.js is installed and that npx runs. The first start downloads the package, so it takes longer than the following ones. Run npx -y @microsoft/powerbi-modeling-mcp@latest --help once in the command prompt, so that the package is already in the cache.

Cannot connect to Power BI Desktop. The file must be open, and the name in the prompt must match the title bar: ContosoDemo10k, not ContosoDemo10k.pbix. After a restart of Power BI Desktop the port changes. Connect again.

The write operations are rejected. You need write permission on the model, and the server must not run with --readonly. See step 8.

The agent reports that a command is not available when it validates the report. Install the two command-line tools of step 11, @microsoft/powerbi-report-authoring-cli and @microsoft/powerbi-desktop-bridge-cli, then restart the client.

The report authoring skill does not modify the report. It requires PBIP in PBIR format. Verify both preview features and save the project again, because enabling PBIR does not convert a project already saved.

The report changes do not appear in Power BI Desktop. Reload the project. Power BI Desktop shows the version loaded before. Verify the preview option of the Power BI Desktop Bridge, so that the agent reloads the project without your intervention.

The agent modifies objects you did not mention. Name the objects explicitly. For example, “add display folders to the measures in the Sales table” instead of “apply the best practices”.

Conclusions

Install Node.js, install Codex CLI, register the Power BI Modeling MCP server with one command, open a model in Power BI Desktop, connect with one prompt.

These are the rules we suggest applying:

  • Trust the folder before you expect the project configuration to be read.
  • Raise startup_timeout_sec when the server is started with npx.
  • Register the server by its package name, instead of asking the agent to find it.
  • Verify the connection with a read-only prompt before any change.
  • Keep a backup of the model, or work on a PBIP project under source control.
  • Read the confirmation prompts.
  • Use --readonly for the first session on a model that matters.
  • Do not write the connection port in a script.
  • Open the folder of the project in the client, so that the model and the report are both visible to the agent.
  • For the report layer, use PBIP with PBIR, and save in Power BI Desktop before the agent iterates.

References

The content of this article was verified in September 2026, with the @microsoft/powerbi-modeling-mcp package installed from npm with the @latest tag. Command names, preview settings, and plan boundaries change over time, so the linked documentation is the authority.

Last update: Sep 08, 2026
In this page