The Microsoft Dataverse plugin for AI coding agents ships eight task-specific skills, from dv-overview to dv-security, that a supported agent loads on demand. You describe the outcome, the plugin routes to the right skill, and the agent chooses between MCP, the Dataverse CLI, the Python SDK, the Power Platform CLI and the Web API. It is a preview feature.
A skill is instructions the agent loads when it needs them
The usual way to make a coding agent good at something is to write a very long prompt and hope it holds. Skills take the other approach. Each one is a focused set of instructions for a kind of task, and the agent pulls in the ones the job needs and leaves the rest alone.
You do not name the skill. You say what you want and the plugin routes the request, often combining several. The source is public, in the microsoft/Dataverse-skills repository on GitHub, so you can read exactly what your agent is being told to do.
The skills, and what each one covers
| Skill | Covers |
|---|---|
dv-overview | Shared rules and tool selection. Loaded before any specialised skill |
dv-connect | Authentication patterns the other skills reuse |
dv-query | Reading data, analysis, notebook work |
dv-metadata | Tables, columns, relationships |
dv-data | Create, update, delete, upsert, import and generate records |
dv-solution | Packaging schema changes and deploying them |
dv-admin | Retention and audit settings |
dv-security | Users, teams and security roles |
Microsoft publishes worked examples of which skills a request pulls in. Connecting and listing tables uses dv-overview, dv-connect and dv-query. Creating a table and loading a CSV file uses dv-overview, dv-metadata and dv-data. Packaging schema changes for a test environment uses dv-overview and dv-solution. Adding an integration user and assigning a role uses dv-overview and dv-security.
The interesting part is the tool routing
Dataverse has five ways in, and choosing badly is how you end up with a job that takes an hour instead of a minute. dv-overview is where the choice is made. It tells the agent to pick by operation and data volume: MCP for interactive queries and small record sets, the Python SDK for bulk operations, transformations, paging and analysis.
For large writes the dv-data skill uses the Dataverse bulk messages CreateMultiple, UpdateMultiple and UpsertMultiple rather than one HTTP request per record, and splits big payloads into batches. That single rule is the difference between a 500-record import that makes 500 calls and one that makes a handful.
It also handles the boring problem that ruins imports. Alternate-key upsert means rerunning the same CSV matches on a natural key such as email rather than creating a second copy of everyone, and multi-table imports are ordered by foreign-key dependency so the lookups resolve.
The rule that keeps it safe is environment first
Before any specialised skill runs, dv-overview makes the agent check whether the workspace is initialised, confirm which environment and which solution it is about to change, and follow an environment-first workflow: make the metadata change in Dataverse, export and unpack the solution, then commit the generated files.
That ordering is deliberate. Hand-editing unpacked solution XML and hoping it imports is a well-known way to lose an afternoon, and an agent is more than capable of generating XML that looks right and is not.
The instruction I like most is the one telling the agent to identify operations that are not covered rather than inventing unsupported commands or SDK methods. An agent that says it cannot do something is worth more than one that confidently makes up a CLI flag, which is the failure mode I keep running into everywhere else.
Where it fits if you already have ALM
It drives your existing process rather than replacing it. dv-solution packages components and deploys them, so the pipeline you already have stays the thing that moves change between environments. If you have no pipeline yet, an agent that can create tables makes that gap more expensive, so fix the ALM first.
Treat the preview label seriously. Preview features carry restricted functionality and supplemental terms, and the skill surface can change under you. Use it on a development environment, watch what it runs, and keep it out of anything scheduled.
Common questions
Is the Dataverse plugin for AI coding agents generally available?
No. It is a preview feature, so it is not meant for production use and the skill surface can change. The skills are published in the microsoft/Dataverse-skills repository on GitHub.
Do I have to tell the agent which Dataverse skill to use?
No. You describe the outcome and the plugin routes the request, usually combining several skills. Creating a table and loading a CSV file, for example, pulls in dv-overview, dv-metadata and dv-data.
Using agents on Power Platform work
Most of the value here is in the routing rules rather than the model. If you want to talk through where an agent helps on your Power Platform build and where it adds risk, book a free 30-minute call. I run this setup daily, including the squad behind this site.