If you run a small company you have probably used Claude or ChatGPT to draft an email or tidy a spreadsheet. What it could not do, until recently, was reach into the software your business runs on and do the work there. MCP is the piece that changes that. It is worth ten minutes of a founder’s attention, because it changes what your software is for, and because switching it on carelessly is an easy way to hand an eager assistant more access than you meant to.
What MCP is, without the jargon
MCP stands for Model Context Protocol. Its own documentation describes it as an open-source standard for connecting AI applications to external systems, and offers a comparison that has stuck: a USB-C port for AI. One standard plug, so any assistant that speaks MCP can connect to any product that offers it, without a custom integration for each pair.
In practice a product publishes an MCP server: a list of actions an assistant is allowed to take, such as “list my tasks” or “create an issue”. You connect your assistant to it, sign in, and from then on a request in plain words is turned into those actions. You are not teaching the assistant your software. The software is telling the assistant what it may do.
Why a small business should care
Large companies have operations staff whose job is keeping the tracker, the timesheets and the HR system up to date. A team of fifteen does not. The work either lands on the founder at eleven at night or does not get done, and then the data is wrong when you need it. An assistant that can do the clerical half changes that sum.
- After a client call, paste your notes and ask for an issue per action point in the project tracker.
- On Friday, dictate the week’s hours against the right projects instead of filling a form.
- Before you approve leave, ask who else is waiting on you and decide in the same conversation.
- When a referral arrives by email, ask for the candidate to be added to the right job.
None of that is clever. It is the typing nobody wants to do, and it is exactly where records fall behind.
The catch: one connector per product
MCP standardises the plug, not your stack. If your projects are in one product, your time tracking in another, HR in a third and files in a fourth, you need four connectors. Each has its own sign-in, its own idea of permissions, its own log, and its own rules about which plan includes it. The people who make them are careful about this themselves: one payroll vendor’s documentation tells customers not to connect other servers in the same session as theirs, because data can pass between them.
That is the less obvious argument for keeping work in fewer places. We have written before about what eight separate subscriptions really cost and about all-in-one versus best-of-breed; connected assistants add a new line to that comparison. One product with one connector means one approval, one log and one off switch.
Five things to check before you switch it on
- Who does the assistant act as? It should act as the person who connected it, with that person’s permissions, not as a shared account with broad access.
- Is reading separate from changing? You should be able to allow questions without allowing edits, tool by tool.
- What is left out entirely? Some things should have no action at all. Passwords and other credentials are the obvious ones.
- Is there a record? Someone will ask “did the assistant do that?”. You need a list of what it did, for whom, and whether it worked.
- Is there one off switch? An owner should be able to stop every assistant in the organisation at once.
The assistants add checks of their own, and they are worth knowing. ChatGPT connects through what OpenAI calls developer mode, lists it for paid and education accounts on the web, and by default asks you to confirm before an action that writes. OpenAI’s own guide calls the mode powerful but dangerous and names the risks plainly. Claude’s help pages note that a custom connector is a service Anthropic has not verified. Both are telling you the same thing: the judgement about what to connect is yours.
How Orbiz does it
Orbiz is one product with ten tools, so it has one connector: Orbiz MCP. An assistant connected through it can take 81 actions across 9 of those tools. That covers issues and sprints in Cadence, timesheets and leave in AMS, tasks and candidates in HRMS, and pay, files, notes and recordings besides. Every member connects with their own sign-in, and the assistant can only do what that member’s role allows.
Against the five checks: it acts as you; reading and changing are separate ticks for each tool, with pay changes ticked separately again; the Vault, where credentials live, has no assistant actions at all and nothing can be deleted; every action is listed in an activity log; and an admin can switch assistants off for the whole organisation. It is included in every plan. The full list of requests it can carry out is in the prompt library.
Where it is not the answer
An assistant is the wrong tool for a decision. It can list the candidates in a round; it should not choose between them. It can tell someone what they were paid; in Orbiz it cannot run payroll, and it cannot clock anyone in or out, because both are records that a person stands behind. And if nobody on the team uses an assistant yet, the connector changes nothing until they do. Start with one person, one tool and reading only, and widen it when it has earned the trust.
Sources, read 4 October 2026: the Model Context Protocol introduction (modelcontextprotocol.io); Claude’s Help Center article on custom connectors; OpenAI’s developer mode guide; Gusto’s MCP documentation.
Built for your industry
