Microsoft 365 Copilot

Who can use this feature?

A maker in your organisation who builds agents in Copilot Studio, working with your N3O contact and your Microsoft 365 administrator.

A Copilot Studio agent is how your team works with your N3O Cloud data from inside Microsoft 365. You build the agent once, in your organisation's own Microsoft 365 tenant, point it at your subscription's address, publish it to your organisation, and your team uses it in Microsoft Teams and in Microsoft 365 Copilot chat. Each person who uses it signs in with their own N3O Cloud account the first time. What the server is and what it can answer is on the MCP Server page.

Before you start

The server has to be turned on for your subscription before any of this works. Ask your N3O contact to enable it, and to give you:

your subscription's address for the server; and
the description to use for it.

The redirect URL is the one thing that goes back the other way. Copilot Studio shows it once the tool is created, and N3O registers it once, before anyone can sign in.

What Copilot Studio asks of your organisation, licences included, is Microsoft's to set and it changes, so read Microsoft's Assign user licenses and manage access before you start.

What it costs

Copilot Studio agents run on Copilot Credits, which your organisation buys from Microsoft. N3O makes no charge for the connection. Without credits in place for the environment, the agent's test chat does not answer, even for a maker who holds a Microsoft 365 Copilot licence, so settle credits before you build. On the standard harness the test chat is not billed; on the GitHub Copilot harness it draws credits.

What a Microsoft 365 Copilot licence covers depends on the agent's harness. On the standard harness, Microsoft makes no charge when someone who holds a licence uses the agent in a Microsoft 365 channel under their own signed-in identity, within the licence's fair-usage limits, and any other use draws credits. On the GitHub Copilot harness, all use draws credits, whatever licence the person holds.

The rates, and what a licence covers, are Microsoft's to set and they change, so read Microsoft's Billing rates and management before you commit to anything. The agent appears in the Power Platform admin centre, under Licensing, once it has billed usage. Microsoft updates usage daily, so allow a day. You can set a monthly limit on what the agent consumes from there. That limit is per agent; there is no limit per person. The largest single influence on the bill is the model you give the agent, which is further down this page.

Your environment

Copilot Studio reaches an MCP server through a Power Platform connector, and the environment you build in sets the policy that connector runs under. Build the agent in an environment set aside for it rather than in your organisation's default environment. Billing for Copilot Credits is turned on per environment rather than per agent, so an environment of its own is what lets your finance and IT teams see what this agent costs separately from everything else your organisation runs in Power Platform. Your Power Platform administrator creates the environment, links the billing to it, and makes sure its policy allows the connector. Choose the environment before you build, because an agent belongs to the environment it was built in.

Everyone who uses the agent has to be in the same Microsoft 365 tenant as the agent. The agent answers with the data your subscription holds, personal data included, so the policy you apply to the connector is the policy you are applying to that data.

The harness

Copilot Studio runs each agent on a harness, chosen when the agent is created and not changed afterwards. A Microsoft 365 Copilot licence covers an agent on the standard harness when people who hold one use it in Microsoft 365 channels, but an agent on the GitHub Copilot harness draws credits for all use, building and testing it included, whatever licence the person holds, as Microsoft's Manage costs for agents powered by the GitHub Copilot harness sets out.

Microsoft's other pages linked here describe agents on the standard harness. Microsoft documents adding an MCP server to an agent on the GitHub Copilot harness as a preview, and its Harnesses in Copilot Studio compares the two.

Generative orchestration

After you create the agent, and before you add the tool, turn on generative orchestration in the agent's settings. The agent cannot use MCP tools without it.

Adding the tool

Add the server to the agent as a Model Context Protocol tool, as Microsoft's Connect your agent to an existing MCP server describes. Enter N3O as the Server name, the description N3O gave you as the Server description, and your subscription's address as the Server URL. Names cannot contain brackets. The agent reads the description to decide when to call the server, so it has to say plainly what the server answers, and it should describe everything the server reaches rather than one part of your data.

Authentication

Under Authentication, select OAuth 2.0, then choose Dynamic (with discovery). Copilot Studio reads the settings it needs from the server itself, so there is nothing to enter, and no client ID or client secret to be given or looked after. Copilot Studio also offers a manual option. It is not the route for this server.

The tool runs as the person asking rather than under one account shared by everyone. That is Copilot Studio's User authentication setting, which sits on the tool itself rather than in this dialog. Check that the tool is set to it, because it is what lets each person sign in to N3O Cloud as themselves.

The agent's own authentication, in its security settings, stays on Authenticate with Microsoft. Microsoft says not to turn off authentication at the agent level when the agent has tools that require user credentials, because those tools then stop working.

Sending N3O the redirect URL

A Redirect URL appears on the confirmation page once the tool is created, and the wizard may label it the callback URL. Copy it and send it to your N3O contact as soon as you have it. Every connector has its own redirect URL, and N3O registers yours once, before a sign-in through the agent can succeed.

If you get as far as creating the connection before N3O has registered it, signing in is turned away with a callback URL mismatch. Nothing is broken and nothing has to be built again. Once N3O confirms, try the connection again and it goes through. A connector built again, in another environment or from scratch in this one, gets a new redirect URL, and N3O registers that one too.

Creating the connection

Once N3O confirms, create the tool's connection and sign in with your own N3O Cloud account. Then ask the agent's test chat a question about your data, and watch the agent call the server and answer. If the test chat does not answer, check that Copilot Credits are in place for the environment before looking for anything else.

Choosing which tools the agent may use

The tool's Tools section lists what the server offers. Check that the tools N3O told you to expect are there, and turn off any your team should not have. If you want the agent to read your data and never change it, turn off the tools that change a record. The individual tools only become editable once Enable all tools is off.

Choosing the model

The model the agent uses is the largest single influence on what it costs. Models tagged Deep are reasoning models, which Microsoft bills per 1,000 tokens on top of the ordinary rate. Choose a model tagged General unless you have a reason not to.

A model from an external provider needs an administrator to allow external models for the environment in the Power Platform admin centre, and to grant people access to that provider's models, before the agent answers them. Microsoft's Select a primary AI model for your agent covers both.

Instructions and knowledge

The agent's instructions decide how it goes about a question. Write instructions that tell the agent to learn what your data holds from the server before it queries anything, rather than working from what a question implies, and to answer only from what the tools return, saying so when they return nothing rather than filling the gap from general knowledge.

Leave out any knowledge source that searches public websites. With one in place, an agent has been seen to take facts from a public website rather than from your records.

Publishing to your organisation

Publish the agent once, then add the Teams and Microsoft 365 Copilot channel and publish again. The channel asks where the agent is available, and that choice cannot be changed after publishing. Then select Submit to org catalog. An administrator approves the request in the Teams admin centre, under Manage apps, and the list of agent requests in the Microsoft 365 admin centre can stay empty. Approval makes the agent available to everyone in your organisation and shares it with them. Microsoft's Connect and configure an agent for Teams and Microsoft 365 covers the channel and the submission.

In an environment created shortly beforehand, the submission can be refused for a while, with a message to check your connection or saying that the Teams channel must be enabled. Nothing needs changing. Wait, then submit again.

Saving is not publishing. A change to the model, the instructions or the tools shows in the test chat straight away, but it reaches Teams only once you publish again, and then only in new conversations. A change to the agent's name, icon or descriptions also has to be submitted for approval again.

Installing the agent for your team

Approval lists the agent in your organisation's store in Teams. Install it for the people who should use it rather than leaving them to find it, from the agent's page under Manage apps in the Teams admin centre, as Microsoft's Preinstall apps for your users describes.

Installing the agent for named people restricts nobody else, because approval leaves it available to everyone in your organisation. To limit who can use it, set its availability in the Teams admin centre to specific users or groups, which can include nested groups, as Microsoft's Use app centric management to manage access to apps describes. Leave the agent's own access in Copilot Studio at everyone, because narrowing it leaves people who installed the agent unable to chat with it.

Each person who uses the agent also needs a security role in the agent's environment that can read connectors, or selecting Connect in the connection manager fails to load the connector's information. The simplest way to give everyone one is a group team for the group the agent is available to, with the Basic User role, which can read connectors, as Microsoft's Manage group teams describes. Nobody then needs a role of their own.

In Teams, people open the agent itself once it is installed, rather than picking it from the To: field of a new chat, which can list entries under the agent's name that do not work.

Your organisation's Teams settings have to allow the agent. Under Custom apps in the Teams admin centre's org-wide app settings, Let users install and use available apps by default and Let users interact with custom apps in preview must both be on, and the Shared Power Apps app must be allowed. Otherwise the agent tells people they have no access to talk to it, or does not respond at all, as Microsoft's Known issues page describes. A change to the org-wide app settings takes a few hours to reach people, and a change to who an app is available to can take up to a day.

First use

The first time someone asks the agent something, it replies with a card asking them to connect first. They select Open connection manager on the card, and in Copilot Studio they select Connect for the N3O connection, sign in with their own N3O Cloud account if asked, and select Submit. Then they go back to the conversation and select Retry on the card. The connection is tied to their own identity, so the agent answers as them. They are asked to connect again when the connection expires or is revoked. Microsoft's Configure user authentication for tools covers creating the connection.

The first time a person's question needs a tool, the agent can also ask their permission to use the connector for that tool, and they select Allow on the prompt. Each person is asked once for each tool, and the answer is remembered. Someone who leaves the card or the prompt unanswered gets no answer to their question, so tell your team to expect both.

Signing in is not by itself enough. The person also needs an account in your N3O Cloud subscription, and has to sign in with that account rather than any other N3O Cloud account they hold. If someone signs in and the agent cannot find them, ask your N3O contact to check their account.