Model.Database must use the built-in script helpers instead.This summary is produced by the author, and not by AI.
[Video placeholder: September release walkthrough.]
This is one of our biggest releases ever. It gives you more control over what your AI agents can do, more visibility into changes to your model, and better support for working alongside external tools. You choose how much to delegate: let an agent inspect the model, run DAX queries, or make changes through C# scripts.
The release also brings improvements to Data Preview, model validation and saving model files, alongside a .NET 10 build.
AI agents can work autonomously for extended periods. That makes reviewing their work more important, especially in semantic modeling, where even being off by a few cents may be unacceptable.
We've added new ways for agents to interact with the model, alongside ways to review and undo their changes if they make a mess.
Individual and corporate comfort levels vary when it comes to delegating work to agents. The MCP server and AI Assistant let you choose what the agent may access and change.
You have three choices to make: where the agent works, which capabilities you give it, and how you review its work.
An MCP server gives AI tools such as GitHub Copilot, Claude Code and Codex access to an application's capabilities through the Model Context Protocol. Tabular Editor's new MCP server connects agents to the copy of the model you have open in TE3.
Connecting the agent to that same session also gives it access to your unsaved edits, so it can work with the model as you currently see it. You can inspect what changed and decide what to keep, adjust or undo.
The MCP server requires the AI features component to be installed. To start the server, go to Tools > MCP Server, choose the permissions for connecting agents and click Start server. Export the configuration for GitHub Copilot CLI, Claude Code, Codex, VS Code agent mode or Cursor, then register the server in that tool. Other clients need to support the server's HTTP connection and authentication. You don't need to configure an AI provider or API key in Tabular Editor for MCP; the connected agent uses its own AI configuration.
There are five permission categories:
Each category can be set to Denied, Read or Write, except Model data, which only allows Denied or Read. MCP permissions are applied when the server starts, so requests don't interrupt you with consent prompts. Restart the server to apply permission changes.
Model metadata: Write allows model edits through C# scripts. Tabular Editor compiles and checks each script before execution. A successful script's model edits form one undo step; a failed script's edits are rolled back automatically. Scripts that fail the safety checks, for example by writing files or sending server commands, open in a C# script tab for you to review and run manually. Opening that tab requires Documents: Write; without it, the request is refused.
You can now let the AI Assistant carry out model changes directly. Enable Allow AI assistant to run C# scripts directly under Tools > Preferences > AI Features > AI Assistant. This requires Model metadata to be set to Write under AI Features > Permissions.
The assistant still uses C# scripts to make these changes. Scripts that pass Tabular Editor's safety checks can run directly; scripts that fail them, for example by writing files or sending server commands, open in a script tab for you to review and run manually. Each direct run is one undo step, and a failed run rolls back its model changes. If you have Preview changes enabled, you can review the changes and cancel to undo them. Asking the assistant to write a script still opens a script tab for you to run yourself.
The assistant also gets a few quality-of-life improvements:
Tabular Editor keeps a copy of the semantic model in memory to track changes, support undo and continuously validate the model. When agents and other editors are also at work, you need to see what changed and decide what to keep. This release adds unsaved-change indicators, selective revert, restore for deleted objects, and automatic reloading of model files changed on disk.
A script or agent can make changes across the model at once. Change indicators show which objects and properties differ from the last saved state, so you can review and, if needed, revert them before saving. The indicators apply whether you, a script or an agent made the changes.
In the TOM Explorer, added objects have a green plus, deleted objects a red minus, and edited objects an orange dot. Their parent tables and folders also show indicators, so you can find changes without expanding every part of the model. Color blindness mode under Tools > Preferences > User Interface > Accessibility uses teal for added objects.
The TOM Explorer and Properties pane each have a Show changes toggle. The TOM Explorer keeps changed objects and the containers needed to reach them visible. The Properties pane highlights changed properties and can filter out the unchanged ones.
Saving clears the indicators. So does returning a changed property or object to its last saved state, whether you edit it back, undo the change or use Revert.
Suppose an agent changed a measure's expression and format string. You want to keep the new expression, but the format string was better before. Right-click Format String in the Properties pane and choose Revert. The format string returns to its last saved value, while the expression stays changed. Edit > Undo brings the edited format string back if you change your mind.
You can also right-click a changed object, its table, a display folder or the Model node in the TOM Explorer and choose Revert. This returns the selected object and everything beneath it to the last saved state: additions are removed, deleted objects come back and edited properties return to their saved values. Changes elsewhere in the model remain. The revert is one undo step. If an object can't be put back, Tabular Editor reports it; the changes it could revert remain reverted.
By default, deleted objects stay visible in the TOM Explorer with a strikethrough until you save the model. Right-click a deleted object and choose Restore to bring it back as it was immediately before deletion, including any unsaved edits.
For example, you edit a saved measure's expression, then delete the measure. Restore brings the measure back with the edited expression, still marked as changed. You can then use Revert on the restored measure to return to the saved expression.
Under Tools > Preferences > TOM Explorer > Unsaved changes, you can choose how long deleted objects remain visible and whether to gather them under a Deleted objects node. Keeping them visible until the model is closed lets you restore them even after saving.
[Video placeholder: Edit a saved measure, delete it, Restore the edited expression, then Revert to the saved expression.]
An agent or another editor might also change the model files on disk. Those edits happen outside TE3's MCP permissions and script checks. Tabular Editor detects changes to the loaded .bim file, or to .json and .tmdl files within the loaded model folder. With no unsaved edits in Tabular Editor, the model reloads automatically.
If you have unsaved edits, the External changes detected prompt lets you choose: Reload discards your local edits and loads the files from disk; Ignore keeps your local edits. The two sets of changes aren't merged.
The preference is named Automatically revert from disk, and is enabled by default under Tools > Preferences > Miscellaneous. It reloads the model from disk; the selective Revert command above returns unsaved edits to their saved state. In workspace mode, a successful reload also redeploys the workspace database. Use your version-control diff to review edits made directly to files.
[Screenshot placeholder: External changes detected prompt with Reload and Ignore visible while TE3 has unsaved edits.]
Tabular Editor keeps a local audit log of AI Assistant and MCP activity, including tool outcomes and the full text of C# scripts run or handed over for review. Use Open audit folder under Tools > Preferences > AI Features to inspect the record. Script text can contain model details, so treat the files accordingly.
In Enterprise Edition, administrators can set limits on what the AI Assistant and MCP server may access. They can also lock AI configuration or restrict which providers and endpoint hosts users may choose. The user-facing permission settings stay within those limits; an administrator can apply stricter limits to MCP than to the in-app assistant.
Organizations can also supply a shared folder of Custom Instructions and disable users' personal instructions. If an organization instruction and a personal instruction share an identifier, the organization's instruction takes precedence. Administrators can withhold individual tools from the MCP server without removing them from the in-app assistant.
Data Preview gets easier to navigate, and you can take the query behind a preview into a DAX Query document for further work.
You can set Preview data to display columns in alphabetical order. Selecting a column in the TOM Explorer also scrolls the preview to that column and highlights it, which is especially useful for wide tables. Column tracking is on by default; alphabetical sorting is off. Both settings are available in Preferences.
[Video placeholder: On the Date table, select a column in TOM Explorer and show Data Preview scrolling to and highlighting it.]
When previewing a filtered and sorted table, you may want to continue working with the result in DAX. Right-click the preview grid and choose Show actual DAX query to open an editable query. The query includes your filters and sort order, without the preview's pagination.
The release also improves Metric View editing and validation, checks more restrictions on user-context-aware columns, and recognizes compatibility level 1706.
Building on the Metric View YAML 1.1 support covered in June's release, you can rename Metric View objects without deleting and recreating them. Importing YAML with duplicate object names reports an error, and expression diagnostics identify whether a problem affects a join, field or measure. Semantic Bridge is available in Enterprise Edition.
Tabular Editor detects restricted uses of user-context-aware calculated columns while you edit. These include references from standard calculated columns, calculated tables and row-level security filters, including references through measures, and using the column as a relationship endpoint. Errors appear in Messages; direct expression references are also flagged in the DAX editor.
The compatibility level determines which features and metadata a semantic model can use. The New Model dialog now offers 1706 for Power BI and Fabric models. The built-in BPA rule for the latest Power BI compatibility level also targets 1706. Existing models aren't upgraded automatically.
Tabular Editor can now use .NET 10. The .NET 10 and .NET 8 builds provide the same Tabular Editor features, so the choice depends on which runtime your machine or organization can use. We recommend the .NET 10 build; the installer will offer to install the .NET 10 Desktop Runtime. The .NET 8 build remains available for machines that can't move to .NET 10.
Existing C# scripts need attention if they call query or command members directly on Model.Database. Those calls bypassed the permission checks used for agent-executed scripts and are no longer available. Replace them with the built-in helpers EvaluateDax(...), ExecuteDax(...), ExecuteReader(...) and ExecuteCommand(...).
For example, replace:
var result = Model.Database.EvaluateDax("1 + 1");
with:
var result = EvaluateDax("1 + 1");
Scripts already using the helpers don't need to change. This applies to both the .NET 8 and .NET 10 builds.
Sales stored in tables/sales.tmdl could previously delete the table's file, leaving the table missing when the model was reopened..platform file. Previously, saving could replace the name with SemanticModel and clear the description. Updating the Fabric workspace from that Git repository would then rename the semantic model and clear its description there too.With this release, we want to make it easier for you to collaborate with AI while staying in command. Choose the permissions and tasks you delegate, inspect what changed, and decide what to keep, revert or adjust. Before you upgrade, check existing C# scripts for the Model.Database query and command calls described above. Let us know how the release works for you through our community forum.
Take your semantic models further with Tabular Editor.
Give Tabular Editor a spin