Connected to Claude in about five minutes.
No code, no API keys, nothing to install on a server. You paste one address into Claude, sign in with your XBert account, and start asking — about a client's books, or about how the practice itself is tracking. ChatGPT and Microsoft Copilot take more setting up, and what your plan and your IT admin have to allow is below. On Claude, most people get their first real answer before the kettle's boiled.
Under the hood: a remote MCP server with OAuth. That's the whole trick.
Full illustrated guide with screenshots: the XBert setup guide →
Before you begin, you need three things.
An XBert account
With your clients' files connected as usual — Xero, QuickBooks, MYOB or FreeAgent, plus Xero Practice Manager if you use it: XPM is what powers the practice-side questions about jobs, WIP and capacity. The MCP works with what's already in XBert. One thing worth checking: the data and reporting tools only run for clients on an AI-enabled plan. A client on any other plan still works for the workflow side — tasks, templates and notes — but ask about its books and the assistant will come back and say the plan doesn't allow it. Not on XBert yet? Start at xbert.io.
Claude, or another AI that speaks MCP
Claude is the straightforward path — web, Desktop or Claude Code, and you paste one address. ChatGPT takes more setting up: custom MCP connectors sit behind developer mode, which is web only, and on a Business, Enterprise or Edu workspace an admin has to switch it on and publish the connector before anyone else can use it. Microsoft Copilot is a different shape again — you reach XBert by building a small agent in Copilot Studio, and someone in IT approves it before the rest of the practice sees it. Each one has its own steps below.
Your normal login
That's the security model in one line: the assistant signs in as you and can only see what your XBert role allows. If you can't see a client in XBert, your assistant can't either. Role isn't the only gate, though — the client's plan decides whether the data tools answer at all. On a workspace plan, connector settings are usually an admin's to switch on — ChatGPT in particular needs one to enable developer mode before you can add anything.
Connect your assistant.
One address, whichever assistant you use. Pick yours below — each set of steps has its own link, so you can send a colleague straight to the right one.
https://mcp-gateway.xbert.io/mcpThe straightforward path. Add it once and XBert is there everywhere you use Claude, with no extra setup on your other devices.
The first button opens Claude’s Add custom connector box with the name and address already filled in — sign in to Claude first and it lands ready to add. Both buttons end up on the same screen, so use the second if you would rather type it in yourself.
- Select Customize in Claude’s left sidebar, then the Connectors tab. Some of Anthropic’s own pages still call this Settings → Connectors, so don’t worry if the wording differs from what you see.
- Select +, then Add custom connector.
- Name it XBert, paste the address above as the remote MCP server URL, and select Add. Leave the Client ID and Client Secret boxes empty — XBert registers itself, so there is nothing for you to fetch from us.
- Select Connect and sign in with your XBert account in the window that opens.
- Back in the chat, XBert now appears in your tools menu. Ask away.
On a Team or Enterprise plan? An Owner goes first.
Members can’t add a custom connector on those plans. An Owner adds XBert once for the whole firm under Organization settings → Connectors — labelled Admin settings in some places — and everyone else then finds XBert under Customize → Connectors and selects Connect. On Free, Pro and Max you add it yourself, with no admin involved — though Free allows only one custom connector at a time.
Step-by-step with screenshots in the XBert setup guide →. Claude’s own version: adding a custom connector →
Claude Code — direct link to this section
Command lineIf you have already added XBert on claude.ai, you may have nothing left to do.
Try this first. Connectors you have added on claude.ai load automatically in Claude Code when you are signed in with that same Claude account. Run /mcp and look for XBert before you add anything. This only applies to a Claude subscription login — not an API key.
If it isn’t there, or you would rather register it locally, run:
claude mcp add --transport http --callback-port 6274 xbert https://mcp-gateway.xbert.io/mcp- In your next session, run /mcp and complete the XBert sign-in when prompted.
- Ask in plain English.
--callback-port 6274 is not optional. XBert registers a fixed port to finish the sign-in, and Claude Code otherwise picks one at random — leave the flag out and the sign-in fails. --scope user makes XBert available from any folder rather than only the one you ran the command in.
If your Claude Code doesn’t recognise --callback-port, update it first — the flag is only in recent versions. Full syntax in Claude Code’s MCP documentation →
ChatGPT — direct link to this section
Developer modeCustom MCP connectors sit behind developer mode, which is web only. Where you find it depends on whether you are on a workspace or a personal plan.
ChatGPT has no one-click install, so the first button only opens the page where you add the connection — you still paste the address yourself. Developer mode has to be on first or there is nowhere to add anything.
On a Business, Enterprise or Edu workspace
- An admin or owner opens Workspace settings and, under Permissions & Roles → Connected Data, turns on custom MCP connectors.
- Once you have access, open Settings → Apps → Advanced Settings and turn on Developer mode.
- Go to Apps → Create, name it XBert with a short description, and paste the address above as the MCP server URL — keep the /mcp on the end.
- Choose OAuth, select Scan tools, sign in to XBert, then select Create. It appears as a draft — an admin or owner publishes it from Workspace settings → Apps → Drafts.
On a personal Pro account
- Open Settings → Security and login and turn on Developer mode.
- Go to Plugins and select +.
- Give it a name and description, paste the address above under Connection, choose OAuth, and sign in to XBert.
Check what your own plan allows. OpenAI’s developer documentation lists Plus as eligible for developer mode, while its help centre lists only Business, Enterprise and Edu — so if you are on Plus and developer mode doesn’t appear, that is why. On Pro, OpenAI documents read-only access, so XBert can answer questions but not take action. Write actions are a beta on the workspace plans. OpenAI has called this area Connectors, Apps and now Plugins, so if the wording differs, look for wherever custom MCP servers are created. OpenAI on developer mode and MCP →
Microsoft Copilot — direct link to this section
Via Copilot StudioCopilot Studio recognises XBert's sign-in on its own — there is no client ID or secret to copy from us. But “Copilot” means five different Microsoft products, and only one of them is a route to XBert.
| Product | Can it reach XBert? |
|---|---|
| Copilot StudioMicrosoft's agent builder. | YesThe one documented route. You add XBert as a tool on an agent, and it recognises our sign-in on its own. |
| Microsoft 365 CopilotCopilot Chat in Teams, Word, Outlook and on the web. | IndirectlyYou can't paste our address in here — Microsoft documents no self-serve path. You reach it by building the Copilot Studio agent above and having an administrator approve it. |
| Create agent / Agent BuilderThe lightweight builder inside Microsoft 365 Copilot. | NoTakes knowledge sources only. Microsoft's own advice is to use Copilot Studio when you need to reach an outside service. |
| Microsoft Copilot (consumer)copilot.microsoft.com, and the Copilot mobile app. | NoOffers a fixed list of connectors — OneDrive, Outlook, Gmail and the like. There is no way to add your own. |
| GitHub CopilotThe coding assistant in VS Code and Visual Studio. | NoA different product entirely. Listed only so nobody goes looking for XBert in the wrong place. |
Microsoft publishes no one-click install link for adding a tool, so this one is click-steps rather than a button. If your Copilot Studio signs you in at copilotstudio.com instead, you are on Microsoft’s newer experience and the path is Build → Tools → Add → Model Context Protocol (MCP).
Three things to sort out first.
- Generative orchestration must be on. In the agent, that is Settings → Generative AI. New agents have it on already. Without it, MCP tools don’t run at all.
- Your IT administrator may need to act. MCP servers in Copilot Studio run over Power Platform connectors, so a data policy that blocks those also blocks XBert — quietly. Send your admin to the Power Platform admin centre (Security → Data and privacy → Data policy), or to Microsoft’s write-up →
- Licences. Whoever builds the agent needs a Copilot Studio or Microsoft 365 Copilot licence. People who only use the finished agent don’t need one. Worth knowing: the free trial can build and test an agent but cannot publish it. Microsoft’s licensing guide →
- Open your agent in Copilot Studio and go to its Tools page.
- Select Add a tool → New tool → Model Context Protocol.
- Server name: XBert. Server description: a sentence or two on what XBert does — Copilot reads this to decide when to call XBert, so it is worth writing properly. Server URL: the address above.
- For authentication choose OAuth 2.0, then Dynamic discovery. You do not need a client ID, a client secret, or any URLs from XBert — that is the whole point of this option. Don’t choose Manual: we don’t issue the credentials it asks for.
- Select Create, then Create a new connection, sign in with your XBert account, and select Add to agent.
Turn most of the tools off. XBert publishes far more tools than Copilot Studio will take — Microsoft caps an agent at 128 and recommends 25 to 30. Switch Allow all off and enable only the ones your firm actually needs, or the agent won’t behave. Once Allow all is off, tools we add later arrive switched off too, so check back after an XBert release. If you’re not sure which to pick, ask us and we will suggest a starting set.
Getting it in front of the rest of the practice.
Publish the agent, then under Channels add Teams and Microsoft 365 Copilot. Expect one more step you can’t skip: the agent isn’t visible to your colleagues until an administrator approves it. Test it in the Test agent panel or in Teams while you wait — sign-in doesn’t work on the Demo Website channel, so that one will look broken when it isn’t. Microsoft on agent approval →
Anything that speaks MCP can point at the same address. Whether it can finish signing in is the part worth checking first.
Goose
Goose is the one desktop assistant besides Claude Code that can complete our sign-in, because Block lets you set the port it comes back on. Set GOOSE_OAUTH_CALLBACK_PORT to 6274 first — the sign-in will not finish without it.
- For anything else: find where your assistant manages connectors or MCP servers, and add a remote MCP server with the address above.
- If it asks for a transport type, choose HTTP (sometimes shown as “streamable HTTP”). If it asks for an OAuth client ID, leave it empty — XBert registers itself.
- Complete the XBert sign-in when your browser opens, then start asking.
Assistants we can’t connect yet, and why.
We only publish a button where we have checked the whole journey — install, sign in, first answer. These ones install fine and then can’t finish signing in, so we’d rather tell you than let you find out.
| Visual Studio Code | Signs in on a fixed port of its own (33418), with no way to change it. |
|---|---|
| Cursor | Signs in on a fixed port of its own (8787), with no way to change it. |
| Windsurf | Its install link only opens servers already listed in Cognition's own registry, so we can't point it at XBert. |
| Zed | No install link and no documented sign-in flow for remote servers. |
| JetBrains IDEs | Documents header-based keys rather than a sign-in flow, so there's no way to reach an XBert token. |
| LM Studio | Same — header-based keys only, no sign-in flow. |
| Cline / Roo Code | No documented sign-in flow for remote servers. |
Using one of these? Tell us which — it helps us decide what to support next.
Your first five prompts.
Start read-only — a couple about a client's books, a couple about the practice itself. Get a feel for how it answers before you ask it to do anything.
Name the client and the period — or the person and the week. “Bayside Cafe, last quarter” beats “that cafe file” every time.
Managing and removing access.
Turn it off in one place.
Remove or disable the XBert connector in your assistant's settings and the connection is gone. Reconnect any time by signing in again.
Access follows your role.
The assistant only ever holds the permissions of the person signed in. Change someone's role in XBert and their assistant's reach changes with it — nothing is granted beyond that.
Everything leaves a trail.
Every write is validated and checked for duplicates first, and each one is marked as a change to confirm, so your assistant stops and asks before it submits. That approval prompt belongs to the assistant rather than to XBert — what XBert guarantees is your permissions, the validation and an audit-log entry on every action.
Sign-in uses OAuth. Your credentials stay with XBert — the assistant never sees your password.
The technical bit, for whoever asks.
Most people never need this. Your IT team might, and it saves a round trip if it's written down.
How the connection works
- Address
- https://mcp-gateway.xbert.io/mcp
- Transport
- Streamable HTTP
- Sign-in
- OAuth 2.1 authorisation code, PKCE (S256) required
- Client registration
- Dynamic — nothing to obtain from XBert
- Scopes
- openid, profile, email, offline_access
XBert is reached over the public internet, so there is no firewall change to make for the assistants that run in the cloud. If your firm manages your devices and filters outbound traffic, ask IT to allow mcp-gateway.xbert.io and auth.xbert.io.
Known limits, stated plainly
- Assistants on your own machine sign in on one fixed port. XBert accepts 6274 and nothing else. Claude Code and Goose let you set it; several others don’t, which is why they’re on the list above rather than behind a button.
- Copilot Studio takes far fewer tools than XBert publishes. Microsoft caps an agent at 128. You choose which XBert tools that agent gets.
- A client’s plan decides what can be answered. The data and reporting tools only run for clients on an AI-enabled plan. Tasks, templates and notes work either way.
- Your assistant does the asking, not us. The prompt you see before a change is submitted belongs to the assistant. What XBert guarantees is your permissions, the validation, and an audit-log entry on every action.
Assistant menus move constantly. The steps above were checked against each vendor’s own documentation on 10 September 2026 — if a menu name has changed since, look for the nearest equivalent, and tell us so we can correct this page.
How to work with it well.
- Start with questions, not changes.
Spend your first week asking — balances, alerts, worklists. Move to drafting tasks and write-backs once the answers have earned your trust.
- Be specific.
Client, period, and what good looks like. Specific asks get specific answers.
- One client before twenty.
Prove a routine on a single file, then let the assistant repeat it across the portfolio.
- Review anything client-facing yourself.
Always. The assistant drafts; you decide what leaves the practice.
- Treat it like a bright new hire.
Clear instructions, check the work early on. Trust builds the same way it does with people — quickly, when the work keeps standing up.