What connectors are
Technically, a connector is an MCP server: a service that speaks the Model Context Protocol, the open standard AI assistants use to call tools and read external data. Claude calls them connectors; ChatGPT first called them connectors and then apps. The principle is the same: the service declares what it can do, the assistant decides when to use it during the conversation, and every call goes through the service's permissions.
A remote connector is connected with a login: the assistant opens the service's page, you sign in with your user and approve the permissions. The password never passes through the assistant, which receives a token limited in time. There are also local servers, which run on your computer and are mainly for developers; for a company's data, remote connectors are almost always the ones used.
How to add one to Claude and ChatGPT
The menus change often, but these are the steps. On Claude (claude.ai and the desktop app):
- Open Settings → Connectors.
- If the service is in the connectors directory, add it from there; otherwise choose “Add custom connector” and paste the address of its MCP server.
- Press “Connect”: the service's page opens, where you sign in and grant the permissions.
- In a chat, turn the connector on from the tools menu and ask a first test question.
Custom connectors need a paid Claude plan; on Team and Enterprise plans an administrator may have to enable them. In Claude Code a remote server is added from the terminal with claude mcp add --transport http, followed by a name and the address, and authenticated with /mcp.
On ChatGPT open Settings → Apps and connectors, turn on developer mode and create a connector, giving the MCP server address and OAuth authentication. Availability depends on the plan and, in business workspaces, on the administrator. For Muvia, the exact steps for each client, including the consent screens, are in the connection guide.
Categories: documents, CRM, data and BI, development
The connectors directory grows every month, and the name matters less than the category: each category answers one kind of question and carries different risks.
| Category | What it brings the assistant | Typical questions | What to watch |
|---|---|---|---|
| Documents and storage | Files, pages, email, calendars | “Summarize the decisions from the last project meeting” | Shared folders that hold more than you think |
| CRM and sales | Customers, opportunities, activities | “Which deals over €50,000 have been stalled for a month?” | Who can edit customer records |
| Data and BI | Datasets, queries, dashboards, reports | “September's margin by business unit, compared with August” | Definitions of the numbers and visible queries |
| Development | Code, tickets, releases | “Which open tickets touch the invoicing module?” | Write permissions on repositories |
| Direct databases | A database's raw tables | Any question, written in SQL by the model | Access to the whole schema, credentials, tables with no shared meaning |
Connecting a database or company data
The temptation is to connect the assistant straight to the ERP database. It works for a demo, but in day-to-day work it creates three problems. The first is definition: the model has to guess which of the many tables holds the real revenue, whether credit notes should be subtracted, which date defines the month. The second is access: a database user usually sees far more than the person asking the question. The third is traceability: who asked what, and with which query, ends up scattered across conversations.
The sturdier route is to connect the assistant to a data layer that is already governed: sources copied from the ERP, saved queries and datasets with stable fields, permissions by project and by role. In Muvia the connected assistant reads the datasets and queries of the projects the person is assigned to, runs read-only SQL queries only if the person has granted the SQL reads permission, and every write is a proposal that a person approves in Lab → To approve. The same question can also be asked inside Muvia to Athena, who saves the query in the project and writes the notebook with the conclusion.
What to check: permissions, where the data goes, revocation
- Permissions: the assistant must work with the connected person's permissions, never more, and the consent page must say what you are granting. Read and write should be kept separate.
- Writes: who approves a change the assistant proposes? A proposal to approve is better than an action already carried out.
- Where the data goes: the service's answers reach the assistant and, from there, its provider, under the terms of its service. Check those terms for the plan you use.
- Administrator control: an administrator must be able to switch connectors on and off for the whole company and choose which assistants are admitted.
- Revocation: each person must be able to revoke their own connections, and an administrator everyone's, with immediate effect.
- Log: every call the assistant makes should be recorded, with the assistant and the connection that made it.
With Muvia the administrator turns connectors on in Settings → AI connectors, and by default the company admits only verified assistants: Claude, ChatGPT and Visual Studio Code. The access token lasts one hour, the refresh 30 days, and a connection unused for 90 days is revoked on its own. The conversation stays with the assistant: Muvia receives only the requests the assistant makes to it.
Data analysis with Claude: example questions
The questions that work best name the project, the period and the comparison. These are requests to try with Muvia connected, replacing the names with those of your projects.
| Question | What the assistant does | Permission |
|---|---|---|
| “Who am I on Muvia?” | Calls whoami: person, company and permissions granted | Read |
| “List the datasets in the Sales project” | Reads the datasets with their columns | Read |
| “The ten customers with the highest revenue in 2025, with the query” | Writes a query and runs it with query_sql | SQL reads |
| “Which dashboards and reports are in the Sales project?” | Lists dashboards and reports | Read |
| “Create a dashboard with monthly orders by region” | Prepares a proposal to approve in Muvia | Write proposals |
For every number, ask for the query: it says which dataset was used, with which filters and what was summed. If the definition is right, the same query run again in Muvia gives the same result; if it is wrong, you see it there and not in the tone of the answer. Long answers come in pages, and it pays to narrow the question to one project and one period.
Frequently asked questions
They are links between Claude and external services, such as document stores, CRMs or data platforms, based on the Model Context Protocol. With a connector, Claude reads the service's data and, if you allow it, acts on it, with your account and your permissions.
Adding a custom connector to Claude requires a paid plan, and on Team and Enterprise plans an administrator may have to enable it. On ChatGPT too, availability depends on the plan and on the workspace settings.
Yes, with an MCP server for databases, but the assistant will have to guess what the tables mean and will work with the database user's permissions. It is more reliable to connect it to a governed data layer, with datasets that have shared definitions and permissions by role, as Muvia does.
The connected service's answers reach the assistant and therefore its provider, under the terms of its service and of the plan in use. Muvia does not read the conversation: it receives only the requests the assistant makes to it.
Remove it from the assistant's settings and, above all, revoke the connection from the service. In Muvia, revocation is done in Settings → AI connectors and applies from the next call.