Orchestration

How to orchestrate Threll from your favourite AI

Threll has an MCP server. That means you can run your voice agents from Claude, ChatGPT, Gemini or Microsoft Copilot – in the same chat you already work in, without writing a line of code.

ByEspen Hareide Nilsen1 Sept 20268 min read

A woman sits at a desk in a small office in afternoon light, typing on a laptop with a blurred chat window on the screen, while a black desk telephone with a coiled cord stands in sharp focus in the foreground on the same desk

Most people who run a voice agent do it from a dashboard. You log in, find the worker, read through a transcript, fix a sentence in the script, enter the numbers to be called tomorrow morning. Nothing wrong with that. It is just one more place you have to remember to go.

Threll has an MCP server. It moves those same operations into the chat you are already in.

Two levels of AI, and they keep getting mixed up

Worth separating clearly, because they live in the same sentence whenever people talk about “AI that calls”.

The voice agent is the one that calls. It speaks Norwegian, Swedish or Danish, follows its script, and does the job in real time while a human is on the line.

The orchestrator is the AI you sit in yourself – Claude, ChatGPT, Gemini or Microsoft Copilot. It calls no one. It does what you used to do with the mouse: lists workers, reads call history, pulls last week's numbers, suggests a change to the script, prepares a round of calls.

“Set it up so Harald calls these five tomorrow at nine, reminding them the contract is expiring” becomes a prepared, confirmed and scheduled round. You wrote one sentence.

The gain is not that it is faster than clicking. The gain is that the orchestration happens in the same place as the rest of your work – in the same conversation where you also have your CRM, your calendar and your inbox.

What MCP is

Think of it as a standard socket.

Before MCP, every vendor had to build its own connection to every AI service: Threll to Claude, Threll to ChatGPT, Threll to Gemini, Threll to Copilot – and the same again for every other vendor, over and over. With MCP we build one connection, and any AI tool that supports the standard can use it. Which is why this is one article rather than one per AI tool.

For you, putting it to use is simpler than it sounds. You paste one address – https://mcp.threll.io – into the AI service you use. You are then sent to Threll, log in the way you always do, pick the account it applies to, and approve what the AI is allowed to do. No API keys, nothing to copy back and forth, nothing you have to renew manually later.

What you approve is bounded. A connection can be set up to read only – call history, transcripts, numbers – or to place calls and make changes as well. The AI never gets more than you said yes to at login, and it cannot help itself to more along the way: if it needs more, you approve again. It also never gets more than you have access to yourself.

What it can do sits close to what you find in the interface:

What you want to do	What the AI gets to see and do
See who exists and how they are set up	Workers, language, phone numbers, active or not
Call out, schedule, cancel	Single calls and rounds of up to a hundred, always with confirmation
Dig into what has happened	Call history, transcripts, events, numbers
Change what the agent says	The call script documents: fetch, propose a new version, take out of use
Keep track of who gets called	People and leads, one at a time or as an import

And one thing you will not find in the interface in the same form: a confirmation before anything rings.

The confirmation is not a formality

Placing a call is the one action the AI can take here that reaches the world outside. So it is blocked by default, and outbound calls happen in two steps.

The first step calls no one. It produces a proposal: which worker, from which number, to whom, when, and with what instructions. The AI shows you the proposal. The call goes out only when you say yes – and an approval can be used only once, so the same round cannot go out twice by accident.

That applies to larger rounds too. If eighty customers are to be called, the whole list is checked up front and approved as one: either all of them go into the queue, or none do. Change your mind before you have said yes, and the proposal is discarded without anything being sent.

Read the proposal. That is the whole reason it is there.

How to connect

Claude

Open Customize → Connectors.
Click + and choose Add custom connector.
Paste https://mcp.threll.io and add it.
Complete the login on Threll's page, pick the account, approve the access.
Turn the connector on in the conversation where you want it.

On Team and Enterprise, members cannot add their own connectors. An owner adds it under Organization settings → Connectors first, and then each member connects with their own login. That split is deliberate: the AI never gets more access than the person using it.

ChatGPT

Custom connectors sit behind developer mode, and that is the part people do not find.

Go to Settings → Apps & Connectors → Advanced settings and turn on Developer mode.
Go back to Connectors and choose Create.
Fill in a name, https://mcp.threll.io, and pick OAuth as the login method.
Acknowledge the notice about custom connectors.
Complete the login to Threll.
In the conversation: Tools → Use apps, and switch it on.

This requires a paid plan. On Business, Enterprise and Edu, an administrator has to allow developer mode for the workspace first – if the toggle is missing, that is usually policy rather than a fault. ChatGPT is stricter than the others about how a server identifies itself, but Threll's server meets the requirements, so there is nothing for you to create manually.

Should you connect an AI – and with what access?

Fifteen minutes setting the connection up together, and deciding what it gets to read and what it gets to do.

Book Meeting +47 24 12 66 30

15 minutes · no obligation · pick your own time

Gemini

Here it matters a great deal which Gemini you mean, because support differs across the three surfaces.

The developer surfaces. Antigravity connects the same way Claude or Cursor does: you point the configuration at the address and approve one login in the browser. Note that Gemini CLI stopped serving free, Pro and Ultra accounts on 18 June 2026. Organisations on Gemini Code Assist Standard or Enterprise, and anyone using paid API keys, still run it.

The Gemini app. Custom connectors go through Spark, the agentic mode, and availability is limited for now. In practice this is not yet a route for rolling something out across a company.

Gemini Enterprise. Here Threll goes in as a Custom MCP Server under connected data sources. This connector does not set itself up: it asks for a few values that have to be created in advance. Get in touch and we will give you what goes in each field.

Or skip the standard. If you are building your own agent on the Gemini API, Threll can hand over the whole tool list preformatted for Gemini, OpenAI or Claude. Then you never have to deal with MCP at all.

Microsoft Copilot

“Copilot” is not one product, and only some of them can connect to an MCP server. Work out which one you are sitting in before you go looking for a button that is not there.

Copilot Studio is the way in. You build an agent there and give it Threll as a tool:

Open your agent and go to Tools.
Choose Add a tool → New tool → Model Context Protocol.
Fill in a server name, a short description and https://mcp.threll.io. The description is not decoration – the orchestrator uses it to decide whether to call the server at all.
Pick OAuth 2.0, then Dynamic discovery as the type. Threll's server registers the client itself, so there is no client ID or secret for you to create.
Select Create, create the connection, and add the tool to the agent.

The agent can then be published to Teams and Microsoft 365 Copilot so the rest of the organisation uses it from there. Two things worth knowing: Copilot Studio supports only Streamable HTTP, which is the transport Threll's server uses, and the connection runs through Power Platform connectors – if you have data policies there, they govern this too.

Microsoft 365 Copilot without Copilot Studio. An administrator can add Threll directly as a custom federated connector under Copilot → Connectors in the admin centre. That route is built for surfacing data rather than taking actions, and it asks for a registration ID from the Teams Developer Portal that has to be created in advance. So it works for reports and lookups in the call history, not for placing calls. Get in touch and we will give you the values that go in.

The Copilot app on your phone or in Windows. It cannot connect to external MCP servers. There is no paste-an-address route there, and it is not something you did wrong.

GitHub Copilot is a separate product with its own configuration file, and behaves like the developer surfaces above.

All of these services move menus and rename things quickly. If the screens look different from the description above, the vendor's current documentation is what counts, not this article.

What it is actually worth using for

The morning summary. “What came in overnight? Which calls were handed to a human, and what were they asking for?” The AI pulls the list, opens the escalated calls, reads the transcripts and summarises. Five minutes instead of twenty in a list of log lines.

Script review. A worker's script is a document, not a field on the worker. The AI can fetch it, read it against last week's calls and propose a new version where the agent actually stumbled. A new version has to finish processing before it is used on calls, so you get a natural pause between “proposed” and “live”.

Call rounds from a list. Import eighty leads as people in one operation, prepare eighty calls in the next, read the proposal, confirm. The scheduler handles the rest.

The questions nobody has answered. Search the transcripts for what callers actually ask about, not what you assumed they would ask about. It is the best raw material for both the script and your website.

Numbers without a dashboard. “How many calls were answered last month, how long were they, and how does that compare with the month before?” Just stay alert to the fact that some of the numbers flatter the agent more than they tell you anything.

And the actual point: several connectors in one conversation. “Read the new leads in the CRM, create them as people in Threll, prepare calls for tomorrow between nine and eleven, and show me the list before anything goes out.” Three systems, one sentence, one confirmation at the end. Nobody built that integration. It is three systems using the same standard socket.

Three things to settle before you hand an AI the phone

Leave the confirmation on. It costs you one keystroke and saves you the one time the AI misreads who “these five” were.

Grant as little as possible. If the connection is only there for reports and summaries, it does not need to be able to place calls. Access is approved at login and cannot be widened without you saying yes again – use that.

Remember the call is an AI conversation regardless of who started it. The disclosure duty in the EU AI Act does not hinge on whether a human or an orchestrator pressed the button. It is cheap to comply with and expensive to ignore. The same goes for the rule for what happens when the agent has to hand over: it does not become less important because the call was scheduled from a chat.

Threll.ai builds voice agents in Norwegian, Swedish and Danish. The MCP server is part of the platform, not an add-on. If someone on the developer side wants the details, they are in the connection guide and the MCP reference.

Frequently asked questions

To whichever AI service you connected. The orchestrator reads transcripts, call history and numbers through your own login, so it never sees more than you can see yourself – but the content it reads is processed by Claude, ChatGPT, Gemini or Copilot under that vendor's terms. Decide that the way you would for any other system you paste customer data into.

The confirmation, which is why it is on by default. The first step calls no one: it produces a proposal naming the worker, the number it calls from, who is being called, when, and with what instructions. Nothing rings until you say yes, and an approval works only once, so the same round cannot go out twice by accident.

Yes, because access is bound to the person and not to the connector. On Claude Team and Enterprise an owner adds the connector once, but each member then completes their own login, and the AI never gets more than that member already has. A connection that exists only for reports does not need permission to place calls – grant it read-only.

Not for Claude, ChatGPT or Antigravity. You paste one address, complete a normal login, pick the account and approve the access – no API keys, nothing to renew by hand. Copilot Studio sits in between: the wizard only asks for the address, but somebody needs the licence and an agent to hang the tool on. Gemini Enterprise and federated connectors in Microsoft 365 Copilot are the exceptions: both ask for values that have to be created in advance, so get in touch for those.

Yes. The disclosure duty in the EU AI Act attaches to the conversation, not to whoever started it, so a call prepared in a chat carries exactly the same obligation as one placed from the dashboard. The same goes for your escalation rules: being scheduled from a chat does not make the handover to a human any less necessary.

Copilot Studio is the practical route: you add Threll's MCP address in the MCP wizard and give your agent the tools. Microsoft 365 Copilot can also take Threll as a federated connector set up from the admin centre, but that route is meant for reading data rather than placing calls. The consumer Copilot app cannot connect to external MCP servers at all. GitHub Copilot is a separate product with its own configuration.

Continue reading more articles