Insights · Blog · Licensing · May 8, 2026
Microsoft introduced the Model Context Protocol for Dynamics 365 in February 2026. That sounds technical, but has direct license consequences for anyone wanting to work with AI agents on enterprise data.
Model-context protocols are a relatively recent idea in the AI world. Anthropic published the original MCP in late 2024 as an open standard, with the idea: AI agents need a unified, secure way to access tools, data, and functions of an enterprise application — without every application having to build its own proprietary interface. Microsoft adapted the standard and introduced it for Dynamics 365 in February 2026 (Licensing Guide May 2026, p. 41).
What that means in practice: if you build an AI agent today — whether in Microsoft Copilot Studio, in Microsoft Foundry, with Anthropic Claude, or with your own LLM stack — and the agent needs to access data in Dynamics 365, it now goes through a standardized protocol. Previously you had to build a custom connector for every data source. With MCP, the agent queries a central interface, called the "MCP server," and gets structured access to what it needs.
This is simultaneously a technical simplification and a licensing question. Microsoft explicitly regulated who pays what and when in the May 2026 Licensing Guide.
Microsoft describes it in the Licensing Guide as follows:
"The MCP (Model Context Protocol) server provides a standardized, secure way for AI agents to access and act on enterprise data through natural language." (p. 41)
Translated into architecture-speak: an MCP server is a central layer between the AI agent and the data system. The agent asks: "Give me all Sales opportunities with value above 100k in Q2." The MCP server checks authentication, validates permissions, runs the query against Dataverse, returns the results structured. The agent doesn't need to know the Dataverse schemas, OData syntax, or security roles — the MCP server does that.
In practice this opens up the architecture. Whoever could previously only build a Sales Insights agent in Copilot Studio (because that was where the Dataverse connectors reliably ran) can now in theory also run an Anthropic Claude agent or a custom LangChain agent against the same MCP server. The only question is — what does it cost?
Microsoft documented a four-tier logic in the Licensing Guide (p. 41):
Direct Microsoft quote: "When MCP tools are utilized within Microsoft Copilot Studio, no additional charges for MCP tool execution will be incurred. (Standard orchestration charges for such agents or tool calls continue to apply as published at billing rates.)"
Practice: anyone building an agent in Copilot Studio that accesses D365 data via MCP pays only the normal Copilot Studio orchestration charges (credits for every agent action). No additional MCP fee. This is the Microsoft-preferred architecture, and Microsoft makes it attractive on price.
"Access to Dynamics 365 data is included with a Dynamics 365 Premium license (Sales Premium, Finance Premium, Supply Chain Premium, and Customer Service Premium)."
Practice: anyone building an Anthropic Claude agent, a Microsoft Foundry agent, or a custom LLM agent outside Copilot Studio that needs to access Dynamics 365 data needs premium licenses for the users whose data the agent sees. The Premium SKUs cover MCP access inclusively.
"Access to non-Dynamics 365 data is included with the Microsoft 365 Copilot USL."
Practice: anyone wanting to access SharePoint, Teams, Outlook, or OneDrive via MCP (classic M365 data) needs a Microsoft 365 Copilot USL license. That's the $30/user/month SKU.
"Access with other applicable licenses will be billed, as published at: MCP tools are billed at the same rate as AI tools (basic), per Copilot Credit consumption rates."
Practice: anyone holding neither Premium licenses nor M365 Copilot USL who still wants to use MCP pays per tool call in credits. That's the most expensive variant — typically unsuitable for productive workloads, acceptable for pilot phases or experiments.
Plus: Microsoft says explicitly that for third-party platforms such as Anthropic Claude or Microsoft Foundry, additional orchestration costs apply at the respective provider, which Microsoft does not cover.
Pricing notice: Microsoft adjusts list prices regularly (currency adjustments, NCE updates, plan restructurings). Figures here are indicative values from May 2026. For current prices including arades CSP conditions, see the License Cost Calculator (licenses.arades.de) ↗ daily.
Three strategic implications we've discussed with clients in the past two months:
Anyone already considering premium licenses — typically for the 1,000 Copilot Credits or the Premium features — now gets MCP access on top, with no upcharge. Anyone planning custom agents in Microsoft Foundry or with their own LLMs should include the Premium upcharge via MCP license inclusion in the calculation. In the Premium post we've covered the economic side in more depth.
Whoever stays in Copilot Studio automatically has MCP without extra cost. Whoever leaves — for Foundry, Anthropic, custom frameworks — pays for MCP either via premium licenses or via credits. That's a noteworthy Microsoft strategy: technically, MCP is offered as an open standard (which promotes mobility), price-wise, Copilot Studio is favored (which protects Microsoft lock-in). We consider this an honest form of lock-in — visible in pricing, not in technical hurdles.
In practice we're seeing setups where clients use Copilot Studio as the main path and add an Anthropic Claude agent for specialized use cases (e.g. complex analytics queries or domain-specific reasoning tasks). That's architecturally sensible — Anthropic Claude delivers better results in some reasoning scenarios than the standard Microsoft LLM stack. But: as soon as the Anthropic agent accesses D365 data, premium licenses are needed for the corresponding users. This is overlooked in many architecture workshops so far.
We at arades use Devonso internally for license-audit workflows and Customer Insights dashboards. It's a web application that today accesses Dataverse, M365, and Entra via classic Microsoft Graph API calls. Over the past weeks we've evaluated whether it makes sense to switch Devonso to MCP.
Three takeaways from this evaluation that we also pass on to clients:
Status as of May 2026: we're leaving Devonso on Microsoft Graph API calls. MCP migration is a candidate for the Q4 2026 backlog when we actually need agent capabilities in production. Correct — we don't need them yet.
If you're planning an AI-agent architecture in which Dynamics 365 data or M365 data plays a role — write to us. We run an architecture workshop where we walk through use case, technical options, and license consequences in parallel. More on our AI advisory offering at AI & Microsoft Copilot. Anyone wanting to think Microsoft-independent from the start — OpenAI directly, Anthropic directly, open-source stacks — finds discussion posts and our advisory profile at Independent Engineering.
30-min initial conversation · free
We walk through use case, technical options, and license consequences in one workshop. Incl. MCP license mapping, Premium upcharge calculation, hybrid strategy.
Related posts
May 8, 2026 · Licensing
What has changed in the D365 license matrix since October 2025. A commented walk-through of the change log.
May 8, 2026 · Licensing
Microsoft has a licensing mechanic we find misconfigured at every third client. Here's how it works.
May 8, 2026 · Licensing
Premium SKUs have included 1,000 credits/user/month each since November 2025. At what usage does it pay off?
May 8, 2026 · Licensing
Team Members is the cheapest D365 license — and the most common source of violations. What the license can and can't do.
