Using Sense from Claude, ChatGPT or Claude Code

An external assistant connection lets you work with your xMatix organization from Claude, ChatGPT, Claude Code and other assistants that speak the open Model Context Protocol (MCP). You ask questions about your records, run your reports and dashboards, search your organization's knowledge and, where your administrator allows it, create and update records, all from inside the assistant. You sign in with your normal xMatix account, and every call runs as you, inside your own permissions. This page covers the user and administrator side. The protocol, scopes and full tool list are on the MCP endpoint page for developers.

Overview

xMatix publishes one MCP endpoint for every organization on the platform. An assistant connects to it, sends you to the xMatix sign-in, and then asks you on an xMatix consent screen which organizations it may use. From then on, the assistant does its own reasoning with its own model, and xMatix gives it data and governed actions. Sense and the Setup Copilot never run on the assistant's behalf, so a connected assistant spends none of your organization's AI credits.

An administrator decides three things for the whole organization: whether assistants may connect at all, whether they may write, and which assistants are approved. Each connection is one person with one assistant, and an administrator can see and disconnect every one of them.

When to use it

You already work in Claude or ChatGPT and want answers from xMatix data without exporting anything.
You want to compare figures across several organizations you belong to in one conversation.
You are an administrator and want to inspect or change setup (data model, layouts, business rules, scripts, workflows, integrations) from Claude Code or another assistant.

For questions inside xMatix itself, use Sense chat. It knows the screen you are on and can act on it, which an external assistant cannot.

How xMatix handles it

Each call passes through these checks, in this order:

Sign-in. The assistant holds an access token issued after you signed in. It lasts 60 minutes, and the assistant renews it in the background. Stays connected access lets it do that without asking you again.
Membership. You must belong to the organization the call targets, or it is refused with "You are not a member of this organization."
Organization state. A locked organization refuses calls with "This organization is currently locked." In a read-only organization, every tool that would change data is hidden.
Organization policy. Allow external assistants must be on, and if there is an Approved assistants list, the assistant must be on it. This is checked on every call and on every renewal, so a change applies at once to existing connections.
Tool visibility. The assistant sees only the tools that satisfy all of these:
- your granted access (Read, Write, Setup);
- the organization's write setting;
- the Sense capabilities granted to any active agent in AI Studio, less capabilities that need a feature licence your organization does not hold;
- for setup tools, your own setup rights.
Your permissions. When a tool runs, your security profiles, record security and restriction rules apply exactly as they do in the product.

Nothing in this chain can widen access. Each layer can only narrow the one before it. Setup changes made through an assistant land in the setup audit trail under your name, exactly as if you had made them on the Setup screens.

The tool that stores integration credentials is never offered to an external assistant. Enter secrets in xMatix itself, so that they never pass through an outside model.

Prerequisites

An xMatix user account that belongs to the organization. No separate seat is checked for the connection itself, but tools tied to a licensed capability appear only when your organization holds that licence. See Licensing.
Allow external assistants switched on by an administrator. It is off by default for every organization.
If the organization keeps an Approved assistants list, the assistant you use must be on it.
The organization's MCP endpoint URL, which your administrator copies from the External assistants page.
For writes: Allow writes switched on, and your own permission to create or edit the records concerned. For setup tools: you must also be an administrator with the matching setup rights.

Procedure

Step 1 — Switch on external assistants (administrator)

Open Setup → Sense AI Studio → Configuration → External assistants.
Select Settings in the page header. The External assistant settings dialog opens.
Turn on Allow external assistants.
Leave Allow writes (create/update records; setup changes for administrators) off for a read-only start, or turn it on.
Optionally choose assistants in Approved assistants (optional). Leave it empty to allow any registered assistant.
Select Save. You see "External assistants can connect to this organization." The header now shows Assistants: Enabled, the write state and the MCP endpoint.

The dialog also lists the connection steps for each assistant, with your organization's exact URL and commands ready to copy. Send these to the people who will connect.

Step 2 — Connect Claude (web or desktop)

In Claude, open Customize → Connectors.
If xMatix is listed, select Connect on it. Otherwise select Add custom connector and paste the MCP endpoint URL.
In the window that opens, sign in with your xMatix account (see Step 5).

Step 3 — Connect ChatGPT

In ChatGPT, open Settings → Connectors → Create (developer mode).
Paste the same MCP endpoint URL and complete the sign-in (see Step 5).

The connector also works in ChatGPT's deep-research mode.

Step 4 — Connect Claude Code

You can use either of two methods:

With the plugin: run /plugin marketplace add xMatixAI/claude-plugin, then /plugin install xmatix@xmatix. Run /mcp to sign in. A sandbox organization uses xmatix-sandbox@xmatix instead. The settings dialog prints the right command for your endpoint.
Without the plugin: run claude mcp add --transport http xmatix followed by the endpoint URL.

Claude Code opens the sign-in in your browser and returns to the terminal when you are done. If you belong to several organizations and want one project always pinned to a single organization, see the MCP endpoint page.

Sign in with your normal xMatix account. If the wrong account is shown, select Not you? Use a different account.
The Allow access? screen names the assistant (marked "AI assistant connector") and lists what it wants to do. For example: "Read the records, reports and dashboards you can already see", "Create and update records on your behalf" and "Stay connected (refresh tokens)".
If you belong to several organizations that allow assistants, tick each one the assistant may use, then pick the default organization it starts in. All are ticked by default. Organizations that have not switched the feature on are not listed. A line tells you how many were left out.
Keep Remember this decision ticked to skip this screen next time (it applies when you have a single organization). Then select Allow. Cancel connects nothing.

Write access appears on the screen only as a request. If your administrator has not allowed writes, the connection is read-only, whatever the assistant asked for. You have 15 minutes to finish the screen. After that it reports that the form has expired, and you start the connection again from the assistant.

Step 6 — Check what the assistant can do

Ask the assistant who it is acting as. It answers with your user, the active organization and your effective access.
Ask something you know the answer to, such as the number of open orders you own.
Ask for something you are not allowed to see. The assistant should report that the record is not accessible, because the query ran as you.

Step 7 — Work across organizations

If you ticked several organizations, you can ask:

"list my organizations" to see the ones the connection covers and which one is active;
"switch to …" to move the current conversation to another organization;
a single question about another organization, which leaves the conversation where it is.

When a connection covers several organizations, each answer is labelled with the organization it came from. To reach an organization you did not tick, reconnect and tick it.

Step 8 — Review and disconnect connections (administrator)

Open Setup → Sense AI Studio → Configuration → External assistants. The Connections table lists every connection in your organization.
To end one, select Disconnect on its row and confirm Disconnect this assistant for this user?. You see "… no longer has access for …", and the row's Status becomes Revoked.
To end every connection at once, open Settings, turn off Allow external assistants and select Save. You see "External assistants are switched off; existing connections were revoked."

Fields

External assistant settings dialog

Field	What it means	Notes
Allow external assistants	Whether any assistant may act as this organization's users	Off by default. Turning it off refuses every call at once and revokes every connection's renewal.
Allow writes (create/update records; setup changes for administrators)	The ceiling for write access	Off by default. When off, every connection is read-only, even if a user approves write access. When on, assistants can create and update records, and administrators can also use the setup tools. Every write still passes the user's permissions and validation.
Approved assistants (optional)	An allow-list of assistant clients	Empty means any registered assistant. The picker offers assistants that have already registered, shown with how they registered. You can also type a client identifier. Anything not listed is refused at sign-in and on every call.

Item	What it shows
Assistants	Enabled or Off
Writes	Allowed or Read-only
Active connections	Connections not yet revoked
MCP endpoint	The URL people paste into their assistant

Connections table

Column	What it means	Notes
Assistant	The assistant's name, with its kind (AI assistant, Integration app or xMatix)
User	The person's display name and email
Access	What was granted: Read, Write, Setup, Stays connected	Write and Setup are highlighted. They show what was granted at connect time, and the write ceiling still applies on every call.
Connected	When the person approved the connection
Last used	When the assistant last renewed its access or reused the remembered approval	This is not updated on every call.
Status	Active or Revoked	Revoked rows stay in the list, and their Disconnect is disabled.

Business rules

Off by default. A new organization allows no external assistant until an administrator turns the switch on.
Switch-off is immediate. The policy is checked on every call, so calls stop at once. Every saved renewal is revoked too, so the person has to connect again after the switch comes back on.
Disconnect takes effect at the next renewal. Disconnect revokes that person's renewal for that assistant in your organization. The access the assistant already holds can last until it expires, which is at most 60 minutes. To stop every assistant at once, turn off Allow external assistants.
A multi-organization connection loses only your organization when you disconnect it. The person keeps the other organizations they ticked.
The write ceiling applies twice. Write and setup access is dropped when access is issued or renewed, and write tools are hidden on every call while writes are not allowed or the organization is read-only.
A remembered approval lasts 90 days. Until then, a person with a single organization reconnects without seeing the consent screen again.
Renewal rotates. Every renewal replaces the previous one. If an old one is reused, the whole connection is revoked and the person signs in again.
Destructive tools (deletes, activations, publishes, imports and similar) are marked as such for the assistant, which asks you before calling them. The assistant enforces that confirmation, not xMatix. Every call still runs within your permissions.

Example

A finance lead at a two-company group wants quick answers about receivables without exporting data.

The administrator opens External assistants, turns on Allow external assistants, leaves Allow writes off and saves. The header shows Assistants: Enabled and Writes: Read-only.
The finance lead adds a custom connector in Claude with the endpoint URL and signs in. On Allow access?, they tick both companies, set the first as default and select Allow.
They ask, "How many open invoices are overdue by more than 30 days?" Claude answers from the default company, and each answer is labelled with its organization.
They ask, "And the same figure for the second company?" Claude answers for the second company, and the conversation stays on the first.
They ask Claude to mark an invoice as disputed. Claude reports that it has no tool that can change records, because the organization is read-only for assistants.
Later the administrator selects Disconnect on the finance lead's row. Within the hour, Claude can no longer renew its access to that organization, and the lead reconnects if they need it again.

Training

Practice exercise

In a sandbox organization:

As an administrator, turn on Allow external assistants with writes off. Copy the MCP endpoint.
As a trainee, connect Claude or Claude Code, sign in and approve the access.
Ask the assistant who it is acting as, then ask for the count of a record type you own.
Ask it to update one of those records.

Expected result: steps 3 answer with the trainee's name and the correct count. Step 4 is refused because no write tool is offered. The administrator then sees the trainee's row in Connections with Access showing Read and Stays connected, and Status Active.

Quick reference

The setting lives in Setup → Sense AI Studio → Configuration → External assistants.
Off by default; Allow writes is a separate switch, also off by default.
Users connect with the MCP endpoint URL from that page and their normal xMatix sign-in.
The assistant sees only what the signed-in person can see.
Several organizations: tick them at sign-in, then ask the assistant to list or switch.
A per-row Disconnect takes effect within the hour. Switching the feature off takes effect at once.
External assistants spend no xMatix AI credits.
Credentials for integrations are entered in xMatix, never through an assistant.

Permissions

To do this	You need
Connect an assistant and read data	A user account in the organization, with the feature switched on. What you can read is governed by your own security profiles and record security.
Create or update records through an assistant	Allow writes on, plus your own create/edit permission on the entity
Use setup tools through an assistant	Allow writes on, plus administrator setup rights for the area concerned
Open External assistants, change its settings, disconnect connections	The Manage AI agents setup capability. Without it, the page is not available and its requests are refused.

Configuration

External assistant settings: the three controls above, under Setup → Sense AI Studio → Configuration → External assistants → Settings.
Which tools the assistant sees: the Sense capabilities granted to your active agents in Sense AI Studio → Agents and Configuration → Capabilities. If no active agent grants a capability, its tools are not offered. See How Sense AI is governed.
What each person can see or change: security profiles, record security and restriction rules, exactly as in the product. See Security model.
Read-only organization: when the organization is placed in read-only state, assistants lose every write tool automatically.

Most organizations run read-only for the first weeks, then turn writes on once the team trusts the results. The rollout guide, with the checks to run before announcing the connector, is Connecting Claude or ChatGPT to xMatix.

Common problems

The assistant says external assistants are not enabled

The organization's Allow external assistants is off, or the assistant is not on Approved assistants. An administrator turns the switch on or adds the assistant. The person then reconnects, because switching off revoked the earlier connection.

The consent screen lists only organizations that allow this assistant. A line under the list says how many organizations you belong to have not enabled external assistants. If none qualify, sign-in ends with a message saying the organization does not allow the application, or the assistant is not on the approved list. If the account belongs to no organization, the message says so. Check that you signed in with the right account (Not you? Use a different account).

The consent screen waited more than 15 minutes, or it was submitted twice. Start the connection again from the assistant.

"You are not a member of this organization"

The signed-in account does not belong to the organization the call targets. Sign in with the account that does, or ask an administrator to add you as a user.

"This organization is currently locked"

The organization is temporarily locked, so no calls are accepted. Try again later.

The assistant cannot create or update anything

The organization has not allowed writes, the organization is read-only, or you lack permission on that entity. If writes were allowed only after you connected, reconnect so that your access includes write access.

The assistant has fewer tools than expected

What it sees is your access, intersected with the organization's capability grants and licences, intersected with your permissions. A read-only connection or organization hides every write tool. A capability that no active agent is granted (reports, dashboards, knowledge) hides its tools. If an administrator changed grants or settings since you connected, reconnect so the assistant reads the current list.

Its renewal was revoked: an administrator disconnected it or switched the feature off, an old renewal was reused, or the connection reached its lifetime limit. Connect again.

Common questions

Do I need to install anything?

No. The endpoint is part of the platform, and every supported assistant has a built-in way to add an MCP connector. Claude Code users can optionally install the xMatix plugin, which saves typing the URL. You need the endpoint URL from your administrator, which is shown on the External assistants page, and your normal xMatix sign-in. No password is ever given to the assistant itself. You sign in on xMatix's own page, and the assistant receives only time-limited access.

Does a connected assistant use our AI credits?

No. The assistant runs its own model, and xMatix only answers its tool calls with data and governed actions. Sense and the Setup Copilot never run on the assistant's behalf, so no AI credits are spent. What the assistant provider does with the conversation is governed by your organization's agreement with that provider.

Can I connect more than one organization?

Yes. If you belong to several organizations that allow assistants, tick each one on the consent screen and pick a default. You can then ask the assistant to list them, switch the conversation to another one, or answer a single question about another one. Each answer is labelled with its organization. An organization you did not tick is unreachable until you reconnect and tick it.

How do I disconnect?

You can remove the connector in the assistant's settings, which stops the assistant using it. The row stays on your administrator's list until it expires or is revoked. An administrator can select Disconnect on the row. That stops the assistant renewing its access, and it loses access within the hour. Turning off Allow external assistants stops every assistant in the organization at once.

Can the assistant see more than I can?

No. Every call runs as you, with your security profiles, record security and restriction rules, and every call is audited under your name. Organization settings and capability grants can only narrow what the assistant sees, never widen it.

The MCP endpoint: protocol, scopes, the full tool list and pinning a project to one organization
Asking Sense about your data
Extending Sense
How Sense AI is governed
Licensing
Sense in Claude and ChatGPT: the product overview