AI assistants · MCP
Claude, ChatGPT or any client that speaks the Model Context Protocol (MCP): once connected, the assistant works in your Custodo organisation with your own permissions. It finds documents and messages, updates tasks and deadlines, and answers with your company’s numbers.
MCP address
https://api.custodo.ai/mcp
One address for every organisation: you choose which one to work in when you authorise the assistant.
For Claude, ChatGPT, Cursor, VS Code, Claude Code, Gemini CLI and other MCP clients.
What it can do
The assistant uses Custodo’s tools under the same rules as the screens: it sees only the modules enabled in your organisation, and does only what your permissions allow.
Finds documents by type, date, counterparty or amount, reads their extracted fields and text, and points to the page each answer comes from.
Searches and reads mail and PEC in your connected mailboxes, and sends messages from them.
Sees and updates tasks, checklists, calendar events and deadlines.
Works the CRM pipeline, contacts, expense reports and time-clock records.
Starts and decides approval flows, prepares signature envelopes and document collections.
Counts and measures what you ask with analytics and your saved dashboards.
Answers as cards
In Claude and ChatGPT, answers about documents, tasks, the agenda, charts and the sales pipeline appear as interactive cards inside the conversation. Every card has «Open in Custodo», which opens the record in the web app, at app.custodo.ai.
A card reads with the assistant’s own permissions and loads nothing from other sites. An assistant that cannot draw cards shows the same answer as text.
How to connect
The address is always https://api.custodo.ai/mcp: only the place each assistant wants it changes.
Once Custodo is listed in Claude’s connector directory, you can also add it from there.
If a menu item does not match, «Connect an assistant» in Custodo carries the steps for ChatGPT.
For a client that only talks to local servers, the ready-made block uses mcp-remote, a bridge that also handles the authorisation.
Authorisation
The assistant never receives your password. The first time it connects, it opens Custodo’s own authorisation page (OAuth 2.1): you sign in with your usual Custodo account and choose two things.
They all start ticked: untick the ones the assistant must not enter.
One choice: «Everything you can do» or «Read only». Read only, the assistant searches and reads, but cannot write, send or delete, whatever it is asked.
You paste no password and no key into the assistant. The connection follows your permissions over time: if an administrator grants you a new one the assistant gets it, and if one is taken away the assistant loses it at once.
Security
A connected assistant opens no side door: it goes through the same permissions and the same logs as the people working in the screens.
The assistant acts as you, in the organisations you chose, with the modules enabled there and the permissions you hold: it cannot do anything you could not do.
Actions that send, delete, cancel or commit something are marked as such, so the assistant asks you to confirm before taking them.
An administrator can turn on «Human confirmation for writing actions»: every change then waits for a person’s approval.
Every change goes into Custodo’s tamper-evident audit log, with the credential that made it. The Integrations call log keeps track of reads too.
You revoke access whenever you like from «Integrations» → «OAuth clients». An administrator also has a switch that turns off assistant access for the whole organisation.
Text in documents and messages written by third parties reaches the assistant as data, never as instructions: an invoice cannot give it orders.
Your data
Custodo sends content to the assistant’s provider — the text of a document, the body of a message, the extracted fields — only when a tool asks for it. The agreement for that processing is between your company and that provider.
What happens to data when you connect an assistant is set out in the privacy notice, in its «Assistenti AI, agenti e integrazioni» section; the terms of use are in the terms and conditions. Both are in Italian.
If something goes wrong
Not answered here? Write to support.
An administrator enables it under «Settings» → «Modules».
The assistant only uses the modules enabled in the organisation it works in. If it still does not see the new tools once the module is enabled, start a new conversation or reconnect Custodo.
An administrator switches it back on in the Integrations settings.
It is the «Module enabled» switch under «Settings» → «Integrations»: off, it stops every assistant, API key and agent in the organisation on its next call. Once it is back on, the assistant works again without being reconnected.
Reconnect it and tick that organisation on the authorisation page.
The assistant only works in the organisations chosen when you authorised it, and only in those you are an active member of. Disconnect Custodo in the assistant and connect it again: the authorisation page asks again where it may go.
It is connected as «Read only», or you do not hold that permission yourself.
To let it write, reconnect it and choose «Everything you can do». If the permission is missing on your own account, ask an administrator: the assistant cannot hold more permissions than you.
Support
For a connection that does not work, a question about permissions and data, or a security report.
Write to [email protected]. If you already use Custodo, name your organisation and the assistant you connected: it helps us understand at once.
Write to [email protected]If you already use Custodo, raise a request from the app’s support: the lifebuoy in every module’s bar, then «New request». Every request gets a code, and the answers arrive there.
Sign in to CustodoThe company behind Custodo is Archiva Srl, VAT number IT03237470236.
Not using Custodo yet? Request a demo
