Teams need a repeatable way to bring Business Central into Copilot Studio without custom middleware. The built-in Business Central MCP Server exposes API pages as tools an agent can call, but success depends on getting the configuration right, staying under tool limits, and keeping governance tight.

This guide walks through creating a Business Central MCP configuration, exporting it for reuse, and connecting it to a Copilot Studio agent by adding the Business Central MCP server as a tool via the MCP onboarding wizard. It focuses on generative orchestration, dynamic tool mode to avoid a large static tool surface, and the limits and Copilot Credits that will shape your design.

Contents

Status check (as of July 31, 2026): Business Central MCP Server (online) is available. Dynamic Tool Mode is available. Generative orchestration is the default for new agents. Copilot Studio enforces tool limits. The Dynamics 365 Business Central connector throttles at 300 API calls per connection per 60 seconds.

What you will build and why MCP beats middleware

In Business Central online, a native MCP Server turns API pages into tools agents can call. Each allowed operation (read, create, modify, delete, and bound actions) becomes a distinct tool. With this approach, there is no custom proxy to maintain and no duplicated business logic. The configuration lives in Business Central, can be exported as JSON, and can be connected directly to an agent.

Compared to building a bespoke API or flow-based facade, the MCP path centralizes authorization and object scope in one place, and it works with generative orchestration so the agent can decide when to call a tool based on its description and the user’s request.

Prerequisites, identities, and DLP guardrails

Before touching configuration, align on governance:

  • Environments and agent policy: Create and use a Dataverse environment where generative orchestration is allowed. Admins can change orchestration defaults per environment.
  • Data loss prevention: MCP servers in Copilot Studio rely on Power Platform connectors. If a data policy blocks the underlying connectors, the MCP server and its tools are blocked for the agent as well.
  • Business Central access model: The MCP Server grants read-only access to all exposed API pages unless write operations are explicitly enabled in its configuration. Decide early which API pages and write operations are in scope.
  • MCP onboarding connection and scope: The MCP onboarding wizard will create a connection for the Business Central MCP server during agent setup. Ensure the identity used to create that connection has the right Business Central permissions for the selected API pages.

Important: API pages of subtype ListPart and CardPart are not supported as MCP tools. Keep your toolset to supported API page types and bound actions.

Create the Business Central MCP configuration

Work in the Business Central web client.

1) Open the MCP configuration surface – Go to the Model Context Protocol (MCP) Server Configurations page. – Create a new configuration and give it a clear, stable name. Configuration names are included in connection inputs later.

2) Scope the surface – Add API pages to expose. For each API page, choose which operations are allowed: read, create, modify, delete, and any bound actions that appear for that API. – Leave write operations off unless there is a clear requirement and a compensating control.

3) Choose discovery and dynamic behavior – Dynamic Tool Mode: On to discover tools at runtime, Off to pre-list tools. See the dedicated section below for trade-offs. – Discover Additional Objects: On if you want the client to discover supported pages beyond the explicit list. Pairing this with Dynamic Tool Mode provides a flexible read-only surface. – Unblock Edit Tools: On only for API pages where write is required, and then choose specific operations per least privilege.

4) Save and review tool count – With Dynamic Tool Mode Off, each operation becomes a distinct tool listed by name. With Dynamic On, discoverable tools will not be pre-listed in Copilot Studio’s tool UI, even though they are available at runtime.

Configuration patterns that work well:

  • Read-only exploration: Dynamic Tool Mode On, Discover Additional Objects On, Unblock Edit Tools Off. This keeps the agent’s write risk at zero while enabling broad discovery.
  • Curated write toolbox: Dynamic Tool Mode Off, Discover Additional Objects Off, Unblock Edit Tools On. Add a small set of API pages and allow only the required write operations.
  • Hybrid: Dynamic Tool Mode On and Discover Additional Objects On for read breadth, plus explicitly added API pages with narrowly scoped write operations.

Important: When Dynamic Tool Mode is On, expect many tools to be discovered at runtime rather than listed in the Copilot Studio UI. This is by design.

Export and version the MCP configuration JSON

Business Central supports exporting and importing MCP configurations as JSON. Use this to promote consistent configurations across sandboxes and production and to make reviews easy.

  • In the MCP Server Configurations page, select your configuration and choose Export.
  • Store the JSON in your repository alongside a short README describing the purpose, the API pages included, and any write operations enabled.
  • To seed another environment, choose Import and select the JSON. Verify the configuration name and confirm any environment-specific values.

This export/import loop enables reproducibility and an approval step before making changes visible to agents.

Connect via the MCP onboarding wizard (Business Central MCP server)

In Copilot Studio, add the Business Central MCP server to the agent as a tool using the MCP onboarding wizard.

1) Open the agent and add the Business Central tool – Go to Tools and select Add a tool. – Choose the Business Central MCP server option in the MCP onboarding flow.

2) Create or select a connection – When prompted, create a new connection or pick an existing one. The onboarding flow will ask for Business Central environment and authentication. Use an identity with appropriate access to the API pages you exposed in the MCP configuration.

3) Provide inputs and bind to your configuration – Set Business Central Environment and Company. – Provide the MCP Server Configuration name you created earlier. If no configuration name is supplied, the server grants read-only access to all exposed API pages.

4) Confirm tool visibility – With Dynamic Tool Mode Off, you will see one tool entry per operation per API page that you enabled. With Dynamic On, expect only system tools to be listed; the agent will discover additional tools at runtime.

At this point, the agent can call Business Central tools via the MCP server tool using generative orchestration.

Dynamic Tool Mode to stay under the tool cap

Copilot Studio enforces a limit on the number of tools an agent can host. Business Central documentation cites an example about a 70-tool surface. MCP servers also count against the tool total for an agent. Large Business Central models easily exceed a static tool cap if every operation on every API page is pre-listed.

Use Dynamic Tool Mode in these cases:

  • The set of API pages is broad and primarily read-only.
  • The number of pre-listed tools would exceed your tenant’s practical agent tool limit.
  • You want to keep the agent flexible as Business Central customizations introduce or rename API pages over time.

Trade-offs:

  • Visibility: Tools discovered at runtime will not appear in the Copilot Studio tool list. This is normal behavior when dynamic discovery is enabled.
  • Predictability: You control scope through the MCP configuration and allowed operations, but discovery may still expand or contract based on API availability.
  • Troubleshooting: When investigating agent behavior, correlate the Business Central API pages and operations that are in scope with what the agent attempted to call rather than relying on the static tool list.

A practical pattern is hybrid: dynamic discovery for read breadth and a small, explicit set of write tools for critical updates.

Generative orchestration: naming and descriptions that help selection

Generative orchestration is the default for new agents. The agent can select topics and tools autonomously based on their names and descriptions. There is no need to wire explicit tool calls into topic nodes for each case.

Design tips for Business Central tools under generative orchestration:

  • Keep the MCP configuration curated: even with generative selection, fewer, clearer tools lead to better outcomes than a sprawling surface.
  • Favor specific API pages for write: expose only the API pages and operations that represent safe, single-purpose actions. This makes tool selection more reliable.
  • Use clear object names: if API pages or bound actions are named ambiguously, selection quality suffers. Align with your Business Central team on object naming to aid discovery.
  • Describe the agent’s purpose clearly: in the agent’s instructions, explain what the agent is for and what Business Central entities it can read or update. This gives the model context when deciding to call a tool.

Throughput and reliability: limits, retries, batching

Numbers in this section are as of July 2026 and can change with licensing and wave releases.

Agent flows and Power Automate limits

  • Action and concurrency limits apply to agent flows the same way they apply to cloud flows. These include caps on actions, parallelism, and debatching behavior. Daily request allocations are enforced per license and performance profile.
  • When using flows behind the agent, align the flow owner’s license and performance profile with expected throughput. Non-licensed application users draw from different request pools.

Business Central connector and API limits

  • Dynamics 365 Business Central connector: 300 API calls per connection per 60 seconds (as of July 2026). Large record payloads above 8 MB are unsupported.
  • Business Central web APIs: Standard protections include rate limiting with HTTP 429 responses when exceeded and request execution timeouts that return HTTP 504 after the limit. Operational OData limits include page size, batch size, and per-environment speed caps. Documented speeds differ between sandbox and production (for example, sandbox around 300 rpm and production around 600 rpm as of July 2026).

Retry and backoff design

  • In flow actions, configure retry policies explicitly. The default is exponential backoff and retry counts vary by performance profile. Set a custom exponential policy for Business Central calls where idempotency is safe.
  • For custom code paths that call Business Central APIs, implement exponential backoff on 429 responses and honor retry-after semantics.
  • Batch and cache reads where possible, and avoid chat patterns that trigger bursts of many small write calls.

A simple guardrail checklist

  • Keep one connection per intended concurrency unit for the Dynamics 365 Business Central connector to avoid tripping per-connection throttles.
  • Use read-only surfaces for exploration to reduce accidental write bursts.
  • Monitor for 429 and 504 responses and reduce concurrency or batch sizes when they appear.

Cost model: estimating Copilot Credits for common tasks

Copilot Studio uses Copilot Credits as the common billing currency for AI tool usage. Text and generative AI invocations on standard LLMs have published rates per 1K tokens.

An illustrative model to anchor budget conversations:

  • Simple lookup with one LLM turn: If a prompt and response together total about 3K tokens, multiply by the published per-1K token rate to estimate credits for that turn.
  • Two-step update (clarify then write): If the conversation totals about 6K to 8K tokens across two turns, apply the same arithmetic for a rough range of credits.

Notes for planning:

  • Use the published AI token rate table to estimate LLM usage, and validate overall consumption with your environment’s capacity reports.
  • Agent flows consume Copilot Studio capacity once converted. Align flow usage with your environment’s capacity model.
  • Choose pay-as-you-go or pre-purchased credit packs based on predictability, then revisit the agent’s prompting to reduce unnecessary tokens.

What to watch carefully

  • Tool limits and MCP servers: MCP servers count against the agent’s tool limit. Keep the number of attached servers small and use Dynamic Tool Mode to avoid an oversized static surface. Business Central documentation cites an example about a 70-tool surface; confirm the current cap in your tenant.
  • Discovery vs. visibility: With Dynamic Tool Mode On, many tools remain invisible in the Copilot Studio UI and are discovered at runtime. This can surprise teams during UAT unless explained upfront.
  • DLP and connector governance: If a data policy blocks the connectors used for MCP access, the Business Central tool is effectively disabled for the agent. Review DLP policies before rollout.
  • Write operations risk: The MCP Server defaults to read-only unless you explicitly enable write. When enabling writes, do so per least privilege and keep a narrow curated set.
  • Throttling and timeouts: The Business Central connector’s per-connection throttle and the web API’s 429/504 behaviors will appear under load. Configure retries and avoid bursty patterns.
  • Capacity and billing: Copilot Credits apply to LLM usage and agent flows consume Copilot Studio capacity. Budget using published rates and then verify with capacity reports rather than assumptions.
  • Release cadence: Copilot Studio and Business Central MCP capabilities continue to evolve. Review What’s new for breaking or enabling changes before each rollout sprint.

Important: Generative orchestration is the assumed mode for new agents. Tool selection depends on clear names and descriptions. If an admin disables generative orchestration for the environment, tool behavior changes and explicit wiring may be required.

My take

The shortest path to a maintainable Business Central agent is a dynamic, read-first surface with a tiny curated set of writes. Add the Business Central MCP server via the MCP onboarding wizard, keep the MCP configuration in one place, and let generative orchestration do the routing. The two levers that make or break production stability are the size of the tool surface and respect for throttling. Keep both small and you will spend your time on outcomes rather than firefighting.

Sources

Categorizado en:

Copilot Studio,

Etiquetado en:

, , , ,