Most people already ask an AI assistant to draft the message, summarise the meeting or tidy the plan. Then they copy the result into the system where the work actually lives: the project board, the timesheet, the leave form. That last step is where the time goes. This guide explains how to let Claude, ChatGPT or Claude Code do it themselves, using the Model Context Protocol, and how to do it without handing an assistant more access than the person using it.
What MCP is, in one paragraph
The Model Context Protocol (MCP) is an open standard for connecting AI applications to other systems. Its own documentation compares it to a USB-C port: one shape of connector, so an assistant does not need a custom integration for every product. A product publishes an MCP server; an assistant that speaks MCP connects to it, discovers the actions it offers, and calls them when you ask for something. Claude, ChatGPT and Claude Code all support remote MCP servers today.
The interesting part for a business is not the protocol. It is the question every connection raises: as whom does the assistant act, and what can it reach?
The access question comes first
An assistant connected with an administrator API key sees everything the key sees. That is fine for a developer poking at a sandbox and wrong for a company where salaries, contracts and personal files live in the same system as the sprint board. The safe default is the opposite: the assistant signs in as one person, through the same login and two-factor check that person uses, and gets exactly that person’s permissions.
The MCP authorisation specification is built for this. It uses OAuth 2.1 with PKCE and requires the client to name the exact server it wants a token for, so a token issued for one server cannot be replayed against another. What a product does with that is up to the product, which is why it is worth asking any vendor three things before connecting:
- Does the assistant act as me, or as an integration user with its own broad access?
- Can I narrow what it may do, tool by tool, when I connect it?
- Is there a record of what it did, and can I or an admin disconnect it instantly?
How Orbiz MCP answers them
Orbiz MCP is the connector for an Orbiz organisation. Each organisation has its own address, and a connection belongs to one member in one organisation. Every request the assistant makes runs through the same permission checks as the web app, so it cannot open a screen your role cannot, or reach a tool your organisation has switched off. When you connect, a consent page lists what the assistant may do, grouped by tool, and you untick anything you would rather keep manual. Changing salaries is never ticked for you.
Some things are not available at all. The Vault has no tools, so an assistant can never read a shared credential. The attendance timer has no tools either: time is recorded by people, not by an assistant on their behalf. Member look-ups return names only, never email addresses, and pay tools never return bank details.
Every action is written to an activity log with the tool, the outcome and the time, without the text of what was written. You see your own log in Settings; admins see the organisation’s, alongside a switch that turns AI assistants off for everyone at once.
Setting it up
Open Settings → Connected AI apps in Orbiz to find your organisation’s address, with a copy button and the exact steps for each assistant. In outline:
- Claude (web, desktop or mobile): Settings → Connectors → Add custom connector, paste the address, press Connect, sign in to Orbiz and approve.
- ChatGPT: create a connector under Settings → Apps & Connectors (developer mode, where your plan requires it) with the address and OAuth, then sign in and approve. ChatGPT’s deep research can then search and read your Orbiz content through the connector’s
searchandfetchactions. - Claude Code: copy the ready-made setup line from the same Settings screen, run it in your terminal, choose Authenticate when Claude Code asks you to sign in, and approve in the browser.
What it can do once connected
The useful requests are the ones that would otherwise take five clicks and a copy-paste. In Cadence, an assistant can file a bug with the steps and environment from your notes, turn a meeting summary into a batch of issues, move work across the board, comment and close. In Trackline it can log hours against a project and milestone or total what you logged this month. In AMS it can show your leave balance and request a day off, and an approver can see and decide the requests waiting on them.
In HRMS it lists the workflow tasks assigned to you, such as a probation review, and hiring managers can add a candidate from an email. In Ledgr it reads your own pay summary and payslips; payroll managers can propose a salary revision for review. Notely, Filespace and Grabr are covered too. The full list, and examples of what people ask, are on the Orbiz MCP page.
Where it is not the answer
An assistant is fast and occasionally wrong. Anything irreversible deserves a person: nothing can be deleted through Orbiz MCP, and a direct salary change needs an explicit confirmation on top of the scope. If your organisation is not ready to let assistants act at all, keep the connection read-only, or leave the organisation-wide switch off until you are. And if the work you want automated lives in a dozen separate products, the connector question multiplies: each needs its own server, its own consent and its own log. That is part of the case for one platform in the first place.
Sources, checked 2026-10-04: the Model Context Protocol introduction and its authorisation specification (modelcontextprotocol.io), Claude’s connector authentication guide (claude.com) and OpenAI’s developer mode and MCP guides (developers.openai.com).
Built for your industry
