[Declarative Agent] MCP dynamic discovery not working

[chull434]

Trying to follow the documentation to get MCP dynamic discovery working but not having any success

Starting from the docs with this sample https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/plugin-manifest-2.4#mcp-execution-spec-object-example-dynamic-discovery gives the following validation message:
Error: Invalid Declarative Agent Document: declarativeAgent.json. Problems discovered: Member run_for_functions in RemoteMCPServerRuntime has 0 items which is less than the 1 required.

Adding "run_for_functions": [ "*"] clears the first validation message but then gets the next validation message BadRequest: Manifest package validations failed. Errors: Declarative Copilot content does not pass our validations; please review and resubmit. ErrorCode: MissingPluginFunction, ErrorDetails: Functions must contain at least one function., Error Mesage,: Functions must contain at least one function.

Adding a single static or fake function clears the validation messages but when chatting to the agent it never sees any other tools. Something just seems broken here or have a missed a step somewhere?

ai-plugin.json schema: v2.4
Declarative agent schema: v1.6
Teams manifest: v1.24
Runtime: RemoteMCPServer

[m365-agent-toolkit-assistant]

👋 Hi @chull434,

Thank you for reporting this issue! We want to let you know that:

⏱️ Our team will personally review your issue
🎯 You can expect a response within 2 business days

✨ This is an AI-generated initial response
🤖 AI-Generated Preliminary Analysis:

Your issue matches a known problem in the Microsoft 365 Agents Toolkit: when using the RemoteMCPServer runtime with dynamic tool discovery, agents fail validation unless at least one static function is declared, and even then, dynamic tools aren’t surfaced as expected.

Root Cause & Workaround:

• The toolkit currently requires a static function in the manifest for validation—dynamic-only (MCP-discovered) tools do not qualify.
• Even with a fake/static function, agents do not “see” MCP tools at runtime due to a client/server negotiation bug: by spec, agent requests must send Accept: application/json, text/event-stream headers. The client omits text/event-stream, causing dynamic discovery to fail.
• Declaring a fake/static function only clears the validation but doesn’t enable discovery of remote tools.

What To Do:

1. Upgrade to the latest Microsoft 365 Agents Toolkit.
2. Ensure your requests include both application/json and text/event-stream in the Accept header. (Recent toolkit versions fix this.)
3. If issues persist, follow GitHub Issue #15421 for updates.

If you still see problems after updating, please provide:

• Your Agents Toolkit version
• Runtime logs during agent execution
• Full content of your manifest and agent JSON definitions

Let me know if you need steps for checking or upgrading your toolkit!

✨We'd love your feedback! Please take a moment to fill out our survey.

---

Note: This is an automated first response generated by AI. A human team member will review your issue and provide a more detailed response soon. We appreciate your patience!

[Alive-Fish]

Thanks for your report @chull434 .

Hi @sebastienlevert and @SuryaMSFT, I have reproduced it and the responses are responded from service. It seems dynamic discovery is not support by service. What do you think?

[sebastienlevert]

Dynamic discovery is not available in the service just yet. It's something we're thinking deeply about, while balancing governance and developer experience. @SuryaMSFT can share more updates. That being said, this doc is wrong and we should remove this part from the walkthrough.

[chull434]

Thanks both for responses and confirming. Would be great to get any roadmap information as this pretty much blocks us due to the nature of our multi tenanted product.

We will probably progress our custom engine agent integration in the mean time, however the lack of support for custom engine agents in outlook and office apps is also a big limitation at the moment, if there was any product time lines on that would be very helpful to us, thanks

[gotenxds]

Hya @mvpkenlin @Alive-Fish, just wanted to add my voice here, we are also facing this issue due to our mcp being multi tenanted, tools are decided after auth - not before.

I understand the need for governance here but I dont think blocking dynamic tools is doing much as it does not stop a malicious actor from changing the implementation of an already approved tool.

Hope you consider working on this as this is a real blocker business wise and not just a developer experience issue.