Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.ntropii.com/llms.txt

Use this file to discover all available pages before exploring further.

Microsoft Copilot Studio is the Power Platform low-code agent builder. It speaks MCP, which means a Copilot Studio agent can call the Ntro MCP server’s tools, fetch its resources, and use its prompts — all from inside a Microsoft 365 / Power Platform tenant. This is useful when you want fund-ops teams who already live in Teams and SharePoint to drive Ntropii without leaving their existing surface.

​
Prerequisites

  • A Microsoft 365 tenant with Copilot Studio licensed
  • Maker access to a Power Platform environment
  • An Ntropii API key with the right tenant scope (API keys)
  • The Ntro MCP server reachable over HTTPS — see transports for the streamable-HTTP setup
Copilot Studio cannot spawn a local stdio child process — it needs an HTTPS endpoint. For a quick proof of concept, run ntro-mcp --transport http --port 8000 locally and tunnel with cloudflared. For production, host the MCP server on a server you control (Azure Container Apps, AKS, App Service).

​
Run the MCP server with HTTPS

# Terminal 1 — start the MCP server
ntro-mcp --transport http --port 8000 --connection production

# Terminal 2 — expose it
cloudflared tunnel --url http://localhost:8000
# → https://random-words.trycloudflare.com
Note the tunnel URL. You’ll register it in Copilot Studio.

​
Register the MCP server in Copilot Studio

1

Open the agent

In Copilot Studio, open (or create) the agent that should have Ntropii access.
2

Add an MCP connection

From the agent’s settings, navigate to Tools → Add tool → Model Context Protocol.
3

Point it at the server

Server URL: paste the HTTPS URL of your ntro-mcp instance. Authentication: Copilot Studio currently supports header-based auth. The Ntro MCP server reads its credentials at startup from --connection / NTRO_API_KEY, so no per-request auth header is needed — leave the auth field as none and ensure the server itself has the right key baked in.
4

Pick which tools to expose

Copilot Studio lets you whitelist specific MCP tools per agent. For a runbook authoring assistant, the typical set is ntro_whoami, ntro_tenant_create, ntro_entity_create, ntro_workflow_create, ntro_workflow_deploy, ntro_workflow_run. Skip ntro_integration_create unless the agent should be able to register data platforms (which usually requires sensitive credentials you’d rather not put through a Copilot prompt).
5

Configure topics or actions

Wire the tools into your agent’s topics. A common starter topic: “Run NAV for $” → calls ntro_workflow_run with the right payload, polls ntro://tasks/{id} for status, and reports back.
6

Test in the Test pane

Type “What tenants do I have access to?” in the test pane. The agent should call ntro_whoami (or fetch ntro://tenants) and reply with a list.

​
Surfacing resources

Copilot Studio agents can use MCP resources as grounding context. Useful URIs:
  • ntro://tenants — let the agent answer “what tenants exist?”
  • ntro://schedule — let the agent answer “what NAV runs are pending?”
  • ntro://tasks/{id} — let the agent answer “what’s the status of run X?”
Configure these in the agent’s Knowledge section so the planner can pull them when needed.

​
Authentication and scoping

The MCP server runs with one API key. That key is tenant-scoped — so the Copilot Studio agent’s surface area is limited to the tenant the key authorises. If you have three tenants, run three MCP server instances (each with its own key) and register them as three separate connections in Copilot Studio. This is the right boundary: a marketing team’s Copilot agent shouldn’t be able to deploy workflows to the operations team’s tenant just because they share an M365 tenant.

​
Limitations to know about

  • No streaming output today. Copilot Studio renders MCP tool results as a single block. Long-running operations (a NAV that takes 10 minutes) need to return quickly with a task ID, then the agent polls ntro://tasks/{id} on a follow-up turn.
  • No interactive HITL signals from inside Copilot Studio. Workflows that pause for human review surface those reviews in the Tenant UI, not in Copilot Studio. The Copilot agent can tell the user “go to the Tenant UI to approve this step”, but it can’t render the approval form itself.
  • Copilot Studio’s MCP support is evolving. If a feature here looks different from what you see in your Copilot Studio UI, check the Microsoft docs for the latest — the integration changed substantially in 2025.

Claude Code

Terminal-based coding agent — same MCP server.

MCP server reference

All transports, tools, resources, and prompts.