Describe the problem/error/question
After the lastest release of a365 commandline tool (1.1.193+79f9a1eaf3) i have been able to use the Microsoft Agent 365 Trigger Node in a workflow.
Using a Anthropic Chat model to power the agent a single “Hello” comsume up 360k tokens per request even though the MCP tooling from Microsoft are turned off. While I expect at most 20k tokens.
Also if I add a System Prompt (You are , you help with…) to try and steer the model the calls to Anthropic fail.
What is the error message (if any)?
System Prompt changes give: only 1 system prompt allowed errors.
Please share your workflow
Share the output returned by the last node
The service is receiving too many requests from you
This request would exceed your organization’s rate limit of 500,000 input tokens per minute (org: d823af68-44e4-476b-b1f2-0de9eb9c4d0a, model: claude-opus-4-6). For details, refer to: Rate limits - Claude API Docs . You can see the response headers for current usage. Reduce the prompt length or the maximum tokens requested, or try again later. View your current limits at Limits | Claude Platform . To raise this limit, purchase credits to advance to the next usage tier at Billing | Claude Platform .
Information on your n8n setup
- n8n version: 2.18.5
- Database (default: SQLite): PostgresSQL
- n8n EXECUTIONS_PROCESS setting (default: own, main): default
- Running n8n via (Docker, npm, n8n cloud, desktop app): docker on Azure Container App
- Operating system: Linux
Hi @slemmen
Your high token usage is being caused by the MCP Client node. Even though you turned off the built-in Microsoft tools, this node is connecting to your own external server and pulling in all the “instructions” for the tools hosted there. Because the AI needs to know how to use every available tool, it attaches all those descriptions to every single message you send. If your server has a lot of tools or very long descriptions, a simple “Hello” becomes a massive request that eats up your token budget and triggers rate limits.
The “only 1 system prompt allowed” error happens because of a conflict between your nodes. Anthropic’s AI is very strict: it will only accept one set of master instructions per request. Currently, you are likely providing instructions in both the Microsoft Agent 365 Trigger node and the Anthropic Chat Model node. When n8n tries to send both, the AI gets confused and rejects the request entirely.
To fix this, first disconnect the MCP Client node and send a test message; if the token usage drops, you’ll know you need to shorten the tool descriptions on your external server. Second, completely clear any instructions or system prompts from the Anthropic Chat Model node, and put all your persona and steering instructions exclusively in the Microsoft Agent 365 Trigger node. This will stop the crashes and bring your costs back down.
Let me know if this is useful.
It reduces the token usage indeed. So I need to work on the MCP server to shorten the descriptions.
The second part is a problem for me, I don’t have the option to clear or change prompts. I can only select the model and some options like max token, temperature, thinking mode etc…
Or is that different in the workflow I shared?
The reason you can’t find a prompt to delete is that the “second” prompt is invisible. To make the agent work, n8n automatically sends a hidden set of instructions to the AI that explains how to use tools and how to format its thoughts. This “internal manual” counts as a system prompt, even though you can’t see it in your settings.
The error happens because Anthropic is extremely strict: it only allows one single system prompt per request. When you type your own instructions into the “System Prompt” field of the Trigger node, n8n is likely sending your text as a separate second prompt instead of merging it into the hidden one. Anthropic sees two different sets of master instructions and rejects the whole request.
Since there is no setting for you to change this, it is likely a bug in how this specific n8n node handles Anthropic models. To get around it for now, try leaving the System Prompt field completely empty to see if the agent works.
You can also file a bug report on Github if you absolutely need custom instructions
I figured that was the case. Because it works when I leave it empty. I’m gonna check if I can give a level of intelligence to the mcp server to allow agents to search for specific tools.