Dynamics 365 MCP cost: Microsoft's fee is the small line on the bill
Dynamics 365 MCP cost for Claude and other non-Microsoft clients is 0.1 Copilot Credits per tool call: €0.00086 at Microsoft's euro pay-as-you-go price, or €0.86 per 1,000 calls. Finance and SCM Premium users pay no per-call fee. The larger cost is Claude's tokens: on Sonnet 5, any tool result above 500 tokens costs more to read.
List prices, 19 September 2026
Who charges what when an agent calls the Dynamics 365 ERP MCP server
| Client and user | Microsoft charge per tool call | Model cost | The caveat in the same row |
|---|---|---|---|
| Claude or another non-Microsoft client, standard user licence | 0.1 Copilot Credits, €0.00086 at the euro pay-as-you-go price | Billed by Anthropic or your cloud at token rates | The tool result the server returns is input tokens on the model side, every time it is read |
| Claude, user with Finance Premium or SCM Premium | None | Billed by Anthropic or your cloud at token rates | The exemption removes Microsoft's fee only; tokens are unchanged |
| Copilot Studio, user without Microsoft 365 Copilot | 5 Copilot Credits per call as an agent action, €0.043 | Included in the agent-action rate | Reasoning models add a premium rate; generative answers are billed separately |
| Copilot Studio, user with Microsoft 365 Copilot | None for employee-facing use under the user's identity | Included | Subject to Microsoft's fair-use limits; only while the agent acts as that licensed user |
| The agent identity itself | No Dynamics 365 user licence with the System agent role | Not applicable | The System agent role grants no permissions; a second role must grant exactly what the task needs |
Euro figures use Microsoft's published pay-as-you-go price of €0.0086 per Copilot Credit. A prepaid pack of 25,000 credits lists at USD 200 a month, which is USD 0.008 per credit if every credit is used.
Less than a tenth of a cent per call is not the number to budget. The number to budget is how many tokens each call hands back to the model.
The driver is the size of the tool result, not the call
Microsoft charges for executing the call. The model provider charges for reading what the call returns, and a model reads every earlier tool result again on each later turn of the same run, because that is how the conversation is sent back to it. The Microsoft fee grows with the number of calls. The token cost grows with the number of calls times the size of what they return, and without caching it grows faster than linearly.
The break-even is short. At the pay-as-you-go rate Microsoft charges USD 0.001 per call. Claude Sonnet 5 lists base input at USD 2 per million tokens, so USD 0.001 buys 500 input tokens. Any tool result longer than 500 tokens costs more to read once than Microsoft charged to produce it, and it will usually be read more than once.
Form tools make this worse by design. Microsoft's documentation says the form tools return the application's view model, the same model the client renders for a person, and that data tools perform standard operations with fewer calls. An agent that clicks through forms makes more calls and receives larger results.
Nobody outside Microsoft has published the size of those view models. We have not measured one either.
One agent run on Claude Sonnet 5 without prompt caching: 20 tool calls, each returning 2,000 tokens. Round inputs, not a measurement; replace them with your own logs.
Microsoft fee
€0.0172
20 calls × 0.1 credits × €0.0086
Tool-result tokens read
420,000
2,000 × (1 + 2 + … + 20): result k is re-read on every later turn
Claude input for those tokens
€0.72
420,000 × USD 2 per million × 0.86
Per 1,000 runs
€17 vs €722
Microsoft fee against Claude input; with 90 % of these tokens read from cache the token side falls to about €137
Prices: Microsoft Learn (0.1 credits per call), Azure Retail Prices API (EUR 0.0086 and USD 0.01 per pay-as-you-go credit), Anthropic pricing (Sonnet 5 input USD 2 per million tokens), all read 19 September 2026. USD converted at 0.86, the ratio implied by Microsoft's own euro and dollar list prices for the same meter. Excludes output tokens, the tool-use system prompt, tool definitions, cache-write premiums and any negotiated discount. The cached figure assumes cache reads at 0.1× base input.
The four things that move the number
Each of these is a design decision made before the agent goes live. None of them is visible on Microsoft's invoice.
- Tool choice: data tools over form tools wherever a data entity covers the operation, set in the agent's instructions, because Microsoft's own documentation says form tools take more calls.
- Named action tools: one class implementing ICustomAPI that returns a short answer replaces a chain of form calls that each return a full view model.
- Prompt caching: cache reads cost a tenth of base input on Claude, so re-reading earlier results is where caching pays back fastest.
- Licence mix: Finance Premium and SCM Premium remove Microsoft's fee entirely; Microsoft 365 Copilot licences make Copilot Studio usage free for employee-facing agents.
Reproduce it
The same arithmetic, with your own run shape
Paste in the number of tool calls per run and the median tool-result size from your agent's logs. The script prints the Microsoft fee, the token side and the break-even result size.
# Dynamics 365 ERP MCP cost for a non-Copilot-Studio client (e.g. Claude).
# List prices read 2026-09-19; replace the run shape with your own logs.
CREDITS_PER_TOOL_CALL = 0.1 # Microsoft Learn, copilot-mcp, "Other agent client"
EUR_PER_CREDIT = 0.0086 # Azure Retail Prices API, pay-as-you-go, EUR
USD_PER_CREDIT = 0.01 # same meter in USD
USD_TO_EUR = EUR_PER_CREDIT / USD_PER_CREDIT # Microsoft's own ratio: 0.86
SONNET5_IN_USD_PER_MTOK = 2.0 # Anthropic pricing, Claude Sonnet 5 base input
calls_per_run = 20 # your number: tool_use blocks per agent run
tokens_per_result = 2000 # your number: median tool_result size in tokens
cache_hit_share = 0.0 # share of re-read context served from cache
ms_fee = calls_per_run * CREDITS_PER_TOOL_CALL * EUR_PER_CREDIT
# Without caching, result k is re-sent on every later turn: t * n(n+1)/2 tokens.
reread = tokens_per_result * calls_per_run * (calls_per_run + 1) / 2
eff = reread * ((1 - cache_hit_share) + cache_hit_share * 0.1)
tokens_eur = eff / 1e6 * SONNET5_IN_USD_PER_MTOK * USD_TO_EUR
breakeven = CREDITS_PER_TOOL_CALL * USD_PER_CREDIT / (SONNET5_IN_USD_PER_MTOK / 1e6)
print(f"Microsoft tool fee per run : EUR {ms_fee:.4f}")
print(f"Tool-result input tokens : {reread:,.0f}")
print(f"Claude input cost per run : EUR {tokens_eur:.4f}")
print(f"Per 1,000 runs : EUR {ms_fee*1000:,.2f} Microsoft vs EUR {tokens_eur*1000:,.2f} tokens")
print(f"Break-even result size : {breakeven:,.0f} tokens per call")With the defaults it prints EUR 17.20 against EUR 722.40 per 1,000 runs and a break-even of 500 tokens. If your median result is under 500 tokens, Microsoft's fee is the larger line and fewer calls is the lever; above it, smaller results are.
When the Microsoft fee is the line to cut
If your agent makes many small data-tool calls, each returning a few rows, the result sizes stay under the break-even and Microsoft's per-call fee becomes the larger share. Then Premium licences or Copilot Studio for Microsoft 365 Copilot users change the bill more than any prompt work. Designing the security role, the named action tools and the credit budget for one agent is what our Dynamics 365 MCP server engagement does before the agent goes live.
Related reading
- Dynamics 365 MCP server: security roles, tools and credit cost — the engagement that designs the role, the action tools and the budget for one agent
- Claude API cost: the three token types that decide the bill — the model side of this calculation in general, beyond Dynamics 365
- Claude cost optimisation — when the token side is already the problem across several agents
- Claude Enterprise blog — the hub for posts on rolling Claude out past the first team
What finance asks before the first agent goes live
Do we pay twice when Claude calls the Dynamics 365 MCP server?
Yes, to two vendors for two different things. Microsoft bills 0.1 Copilot Credits for executing each tool call on its server. Anthropic, or the cloud you run Claude on, bills the tokens the model reads and writes, including every tool result. In Copilot Studio the two are folded into one agent-action rate of 5 credits per call instead.
Who is exempt from the per-call charge?
Users with Dynamics 365 Finance Premium or Supply Chain Management Premium licences, for agents built on clients other than Copilot Studio. The exemption covers Microsoft's tool execution fee only. The model's token bill is unaffected, and so is the requirement that people chatting with the agent hold a Dynamics 365 user licence.
Does Business Central charge per MCP tool call too?
Microsoft's Business Central MCP overview does not say, as of 19 September 2026. The server is a public preview, read-only by default, and the page we read documents endpoints, headers and authentication but no per-call billing for non-Microsoft clients. Ask Microsoft or your partner before planning volume; do not assume the Finance and Operations rate applies.
When is Copilot Studio cheaper than Claude on the same server?
When the users already hold Microsoft 365 Copilot licences, because employee use is then included and neither the 5-credit agent action nor any token bill applies. It can also win on long runs with large results, because its agent-action rate does not grow with context, unless a reasoning model adds its premium rate. Claude wins when the agent must also read systems Microsoft does not host. Our blueprint, fixed at €4,900, prices both routes for one agent.
What does this calculation leave out?
Three things we could not source or did not measure. The size of a real Dynamics 365 form view model in tokens, which is why the worked example uses round inputs and the script measures yours. Output tokens, cache-write premiums and reasoning tokens, which all add to the model side. And negotiated Microsoft or Anthropic discounts, which only your contracts show.
Next step
Get the cost of one agent written down before it runs
Two weeks, fixed price: one Dynamics 365 agent task, the security role it runs under, the tools it may call and a monthly estimate for both the Microsoft and the model side.