ARTICLE · DYNAMICS 365 · AI AGENTS

What is an MCP server? Dynamics 365 and Dataverse data for AI agents (Copilot, Claude)

MCP (Model Context Protocol) is an open standard that lets an AI assistant – Copilot, Claude or your own agent – reach company data and systems in a uniform way. Microsoft Dataverse, the data platform behind the Dynamics 365 CRM apps, can itself act as an MCP server: once enabled, an AI agent can search, query, create and update records in it, always with the signed-in user’s permissions. This article explains what MCP is, what the Dataverse MCP server can do, how to enable it, what it costs and what to watch for on security.

In short

  • MCP: a standard open-sourced by Anthropic on 25 November 2024; client–server design: the AI application is the client, the data source is the MCP server.
  • Dataverse MCP server: Microsoft’s official MCP server for Dataverse (`https://<org>.crm.dynamics.com/api/mcp`); it searches, queries, manages records and tables and calls Custom APIs.
  • Clients: Copilot Studio agents (enabled by default), GitHub Copilot in VS Code and the Copilot CLI, Claude desktop and Claude Code (an administrator must enable these).
  • Cost: since 15 December 2025 it is charged when an agent built outside Copilot Studio calls it; with qualifying Dynamics 365 Premium or Microsoft 365 Copilot licences, access to Dynamics 365 data is not.
  • Security: Dataverse security roles and row-level security apply; the AI sees only what the user can see.

Author: Norbert Madarász, Managing Director and Partner · Updated: 15 November 2026

What is MCP (Model Context Protocol)?

Anthropic open-sourced MCP on 25 November 2024 to connect AI assistants to the systems where data lives through one standard protocol, replacing fragmented integrations written separately for each system. The design is simple: developers either expose their data through MCP servers or build AI applications (MCP clients) that connect to those servers.

MCP clientThe AI application the user works in: for example a Copilot Studio agent, GitHub Copilot in VS Code, Claude desktop or Claude Code.
MCP serverThe service offering “tools” to the client: search, query, create a record and so on. The Dataverse MCP server is one.
ToolA well-described operation with input and output. The model decides when to call it; the client can control which tools it may call.

The practical difference from a plain chat: with MCP the AI does not just talk about what the model “knows” – it works with the company’s current data and, if allowed, changes it.

What can Microsoft’s Dataverse MCP server do?

According to Microsoft Learn, Dataverse as an MCP server provides intelligent access to tables and records for various MCP clients. The generally available tool set is on the `/api/mcp` endpoint; the `/api/mcp_preview` endpoint adds preview tools if an administrator turns them on.

Tool groupDataverse MCP toolsWhat it is for
Search, readsearch_data, search, describe, read_querysearch data and schema, describe tables and records, Dataverse SQL SELECT queries
Recordscreate_record, update_record, delete_recordcreate and update rows; delete only after explicit user approval
Schemacreate_table, update_table, delete_tablecreate and change tables; delete only after explicit approval
Filesinit_file_upload, commit_file_upload, file_downloadupload and download files via a temporary (SAS) link
Business logicinvoke_api, upsert_skill, delete_skillcall your own Dataverse Custom API, manage agent skills/playbooks
In 2026 Microsoft renamed the tools: describe replaces describe_table, list_tables and fetch, data search is now search_data, and search now searches metadata. If you allow or deny tools by name in your client, review your list. search_data only appears when Dataverse search is enabled.

How to enable it, and which AI works with it?

  1. Power Platform admin center: Environments → the environment → Settings → Product → Features → Dataverse Model Context Protocol: “Allow MCP clients to interact with Dataverse MCP server” must be on. This needs the Power Platform administrator role.
  2. Allowed clients: Advanced Settings lists the clients. Copilot Studio is enabled by default; the others (GitHub Copilot, the Dataverse CLI for Claude etc.) must each be set to “Is Enabled = Yes”.
  3. Copilot Studio: add the Dataverse MCP server to the agent as a tool; for several environments the Copilot Studio MCP onboarding wizard helps.
  4. Claude desktop, Claude Code: either via the local proxy in Microsoft’s `@microsoft/dataverse` npm package (one-time tenant admin consent for the Dataverse CLI app is required) or directly to the remote endpoint by registering your own Microsoft Entra app.

It can be switched per environment, and one client can hold several environments. Turning it off stops every tool and agent that relies on it.

What does Dataverse MCP cost?

  • Agent built in Copilot Studio: per Microsoft’s FAQ the charge applies to agents created outside Copilot Studio; a Copilot Studio agent consumes its own Copilot Credits.
  • Other agents (e.g. Claude, GitHub Copilot): since 15 December 2025 Dataverse MCP tool calls are charged. Exception: with qualifying Dynamics 365 Premium licences or a Microsoft 365 Copilot user licence, access to Dynamics 365 data is not charged.
  • Metering: search_data in usage-based Copilot Credits; all other tools, including metadata search, at the “basic” text and generative AI rate per 10 responses.

The exact credit rate is in Microsoft’s current rate card; Copilot Studio credit packs are covered in Copilot Studio agents for Dynamics 365. We supply the licences from one source: Microsoft 365 and Dynamics 365 licensing.

Is an MCP server secure?

The Dataverse MCP server does not bypass the Dataverse security model: according to Microsoft it respects security roles and row-level security, and users can only reach the tables and records their role permits. Authentication uses Microsoft Entra ID. What still needs attention:

  • Who sees what: the AI sees as much as the user. If roles are too broad, the AI exploits that more efficiently – clean up permissions before MCP.
  • Write and delete: delete_record and delete_table only run after explicit user approval, and every write tool can also be disabled in the client. Note: Microsoft’s sample agent instructions explicitly say not to ask for confirmation on deletes – do not copy that into production.
  • Which client: administrators enable clients per environment; a client that is not enabled cannot connect.
  • Where data goes: retrieved records are processed by the chosen AI provider (Microsoft, Anthropic etc.); review contractual and data-processing terms before go-live.
  • Environments: start in a development or test environment without real customer data; see test data management.

Custom MCP server or Microsoft’s?

AspectMicrosoft’s Dataverse MCP serverCustom MCP server
Good forgeneral data work: search, queries, managing records and tablesnarrow business operations (e.g. “create a quote”), several systems together, your own rules
Developmentnone, just enabling and permissionsrequired: code, hosting, versioning
Controlthe full tool set, narrowed by client-side tool restrictionsonly the tools you write exist
Business logicyour own Custom API can be called via invoke_apianything, including systems outside Dataverse
CostMicrosoft’s billing rule (see above)hosting and maintenance, plus the APIs it calls

Our rule of thumb: as long as the job is general data work in Dynamics 365, Microsoft’s server is enough, narrowed by client-side tool restrictions. If the agent must carry out a business operation (say, assembling a quote or coordinating several systems), we put it in a Dataverse Custom API or a custom MCP server, so the rule lives in code, not in the prompt.

MCP connector, Copilot connector, ChatGPT and Claude connectors: what is the difference?

The word “connector” is now used for at least five different things, and each moves data between Dynamics 365 and AI in a different way. The key difference: either the AI calls the data source live (that is MCP), or the data is indexed in advance into a search index (the classic Copilot connector).

  • MCP connector in Copilot Studio: an MCP server is added to the agent as a tool (Tools → New tool → Model Context Protocol wizard). Every tool and resource the server publishes appears automatically, and Copilot Studio follows changes on the server. Generative orchestration is required; Copilot Studio supports the Streamable transport and has not supported SSE since August 2025. This is also how the Dataverse MCP server gets into a Copilot Studio agent.
  • Microsoft 365 Copilot connector (formerly Microsoft Graph connector): brings external data into Microsoft 365 Copilot and Microsoft Search. The *synced* type indexes the data into Microsoft Graph; the *federated* type queries it live via MCP, without indexing. Microsoft offers more than 100 prebuilt connectors, Dynamics 365 among them. According to Microsoft Learn, federated connectors are currently read-only.
  • Power Platform custom connector: a wrapper around a REST or SOAP API with triggers and actions, for Power Automate, Power Apps, Logic Apps and Copilot Studio. It is not MCP, but in Copilot Studio an MCP server can also be added as a custom connector.
  • ChatGPT: in ChatGPT on the web, a custom MCP server can be added as a plugin, with read and write tools. Write actions ask for confirmation by default, and workspace permissions apply.
  • Claude connector: on claude.ai, Claude Desktop and Cowork, a remote MCP server can be added as a custom connector. The connection starts from Anthropic’s cloud, so the server must be reachable from the public internet on HTTPS port 443. On Team and Enterprise plans an owner adds the connector, and users connect individually, with their own permissions. In the Anthropic API, “MCP connector” is the name of the feature that lets the API call a remote MCP server.
OptionHow the AI gets the dataRead or writeWho enables itWith Dynamics 365
MCP server in Copilot Studiolive tool callboth, depending on the server’s toolsthe agent maker; for Dataverse MCP also the Power Platform adminthe Dataverse MCP server as a tool
Microsoft 365 Copilot connector, synceddata indexed in advance into Microsoft Graphreadthe Microsoft 365 adminprebuilt Dynamics 365 connector for Copilot Chat and search
Microsoft 365 Copilot connector, federatedlive query via MCP, no indexingcurrently read-onlyadmin enables, user authenticateslive, sensitive data
Power Platform custom connectorAPI call via trigger or actionboththe makers and admins of the Power Platform environmentPower Automate flows, Power Apps apps
ChatGPT or Claude connectorlive tool call to a remote MCP serverboth, with confirmationthe workspace or organization admin; for Dataverse also the Power Platform adminthe Dataverse MCP server, if the client is allowed

What do we recommend for Dynamics 365 data?

  • If your people work in Microsoft 365 Copilot and want to search and summarize: the Dynamics 365 Copilot connector is the simplest option, set up by an admin without development.
  • If the agent has to create or update records: a Copilot Studio agent with the Dataverse MCP server, or a custom MCP server for narrow business operations.
  • If the company uses Claude or ChatGPT: the Dataverse MCP server can be connected as a non-Microsoft client, but the admin has to allow the client per environment, and calls may be billed (see above). Microsoft documents Claude Desktop and Claude Code; with other clients we test in a sandbox environment first.
  • For a fixed, repeating process (for example “new lead → task for the area rep”): no AI is needed; a Power Automate flow with a custom or prebuilt connector is cheaper and more predictable.

How we help

We design and run Dynamics 365 and Dataverse environments, so we know the security roles, environments and Custom APIs that MCP builds on. We set up the Dataverse MCP server and the clients, write the Custom API or custom MCP server needed, and test the agent. Details: AI agents and MCP for Dynamics 365, Dynamics 365 consulting, Dynamics 365 support; Dataverse basics: Microsoft Dataverse.

Frequently asked questions

What is an MCP server, simply put?

A service that exposes a system’s data and operations to an AI assistant (Copilot, Claude) as standard “tools”. Anthropic made MCP an open standard in November 2024.

What does the Dataverse MCP server do?

It makes Dataverse, the data platform of the Dynamics 365 CRM apps, available to AI agents: it searches data and schema, runs Dataverse SQL queries, creates or updates records and tables, handles files and calls Custom APIs.

Are MCP servers secure?

The Dataverse MCP server enforces Dataverse security roles and row-level security, with Microsoft Entra authentication; administrators enable clients per environment. The risk lies in overly broad permissions and write tools, so these must be narrowed.

Can Dataverse MCP be used with Claude?

Yes. Microsoft Learn documents Claude desktop and Claude Code setup via the @microsoft/dataverse local proxy or your own Entra app; an administrator must first enable the client in the Power Platform admin center.

What does Dataverse MCP cost?

Since 15 December 2025, calls from agents built outside Copilot Studio are charged in Copilot Credits. With qualifying Dynamics 365 Premium or Microsoft 365 Copilot licences, access to Dynamics 365 data is not charged.

How do you build an MCP server?

Using MCP’s open specification and SDKs: the server defines tools with inputs and outputs and calls the underlying system with authentication. With Dynamics 365 it is often simpler to put the logic into a Dataverse Custom API and call it via the invoke_api tool of Microsoft’s server.

Can the Dataverse MCP server be turned off?

Yes, per environment in the Power Platform admin center. Turning it off stops every tool and agent that relies on it.

What is an MCP connector?

A connector that links an AI assistant or agent to an MCP server, so the AI can call the server’s tools live. In Copilot Studio a wizard adds the MCP server to the agent’s tools; in Claude a remote MCP server can be added as a custom connector, in ChatGPT as a plugin.

What is the difference between a Copilot connector and MCP?

A synced Microsoft 365 Copilot connector (formerly Microsoft Graph connector) indexes external data into Microsoft Graph in advance, so Copilot and search can find it. With MCP the AI calls the source system live and can also write, if the server offers such a tool. Federated Copilot connectors already query live via MCP, but according to Microsoft Learn they are currently read-only.

Can ChatGPT or Claude be connected to Dynamics 365?

Yes, through MCP: the Dataverse MCP server can be used with non-Microsoft clients if the Power Platform admin allows the client in the environment. Microsoft documents the setup for Claude Desktop and Claude Code; ChatGPT can add a custom MCP server as a plugin. For agents built outside Copilot Studio, Dataverse MCP calls may be billed.

Let’s talk about your project.

Tell us what you are planning and we will get back to you within a few days.

Get in touch

Cookie settings

Choose what you consent to. We remember your choice for 6 months; change it any time via “Cookie settings” in the footer.

Privacy notice