Skip to content
Open the app

Connect an AI assistant to the console (MCP)

The console runs its own MCPAn open standard that lets AI assistants call tools on a server in a consistent way. server, separate from the public Knowledge Base one. A staff member connects an assistant they already use, from the ones their institution has approved, signs in with their university account, and the assistant can then work in the console as them: reading the cohort's weekly content and writing drafts, and, where the institution allows it, reading the same aggregate figures the Insights page shows.

Available to every institution, off until you switch it on

First Six has made it available to every institution since 18 October 2026. It stays off at yours until someone who manages integrations approves the assistants staff may connect and switches access on, and staff can connect only those assistants. The data processing agreement (clause 3.7), the pilot charter and the privacy policy describe it.

What it can and cannot do

It acts as the person who connected it, with their permissions and nothing more.

It canIt cannot
List cohorts, audience groups and help categoriesPublish anything
Read a week's existing content and your university's connected source pagesChange settings, roles or people
Draft events, weekly blocks, links, help routes and Ask Anything answersRead or change an individual student's information
With the insights switch: read exact check-in figures and what's weighing on students, as group figuresTouch help requests or the crisis pathway

Everything it writes is a draft. A person reviews and publishes it in the console, exactly as with the console's built-in assistant, and every call it makes is written to your institution's audit log with the assistant's name, the tool it used, and the draft it created. A draft aimed at a group that does not exist is refused rather than saved for everyone, and the assistant is told who each draft will reach. A help route can be drafted only under one of your institution's live help categories: the assistant reads their keys from the list, and a category that does not exist or has been retired is refused before anything is saved.

The insights figures are the same exact figures the staff member sees in Insights, at every group size and within their own scope, so a small group's figures can describe one or two students. A cohort outside their scope gets an answer that says so, not an empty report (for the weekly pulse, which cannot tell the two apart, the answer names scope and no check-ins yet as the possible reasons). The figures are aggregates only: no names, ids, contact details, anything a student wrote, or reasons given before 18 October 2026. Each answer is built field by field from a fixed list, so a column the database gains later cannot slip through.

Three tools carry them: the daily check-in overview (with what was weighing on students who had a wobbly day), the weekly check-in pulse, and What's weighing on students, the reasons students picked after a wobbling weekly check-in. Reasons are sent only for answers given after the student was told, on that screen, that their support team would see them. Reasons given before 18 October 2026 were collected on the promise of group totals only, so they are never sent, not even their categories, and the assistant is told they are left out. The evidence pack is not offered: it stays in the console, where its reporting window is fixed.

Three switches and an approved list, off and empty by default

  1. First Six makes it available. A platform setting, on for every institution since 18 October 2026.
  2. Your institution chooses the assistants, then switches it on, in Settings, Integrations, under AI assistants, by someone with the manage integrations permission. Content access and insights access are separate switches, and neither can be turned on until at least one assistant is approved. Turning insights access on also asks them to confirm that your institution directs it, because the figures go to the assistant's provider, which may process them outside Australia; it cannot be turned on without that confirmation, and the audit log records who gave it. Turning either switch off disconnects every assistant straight away.
  3. Each person consents on a consent screen when they connect, and needs the author content permission for content access or the view insights permission for insights access.

The switches and the approved list are re-checked on every request, so a permission removed this morning stops working this morning.

Which assistants staff may connect

Staff can connect only the assistants your institution has approved, and with nothing approved nobody can connect anything. An assistant is recognised by the exact address it sends people back to after they sign in, never by the name it gives itself: anyone can call their software "Claude", but only Claude's own servers sit behind Claude's address. You can approve:

  • Claude (Anthropic) or ChatGPT (OpenAI), which First Six recognises at their exact return addresses.
  • Apps on the staff member's own computer, such as a desktop or command-line tool that signs in through an address on the same machine. First Six cannot tell which app it is, so this is never approved for you. Approve it only if each staff member uses a computer of their own: on a shared or virtual desktop (a computer lab, a remote desktop or a terminal server), software someone else is running on the same machine could receive the connection.
  • Another assistant by its exact return address, such as your institution's own gateway, under a name your staff will recognise. Staff see it as your institution's name for it, never as recognised by First Six. An address on a host First Six recognises (such as claude.ai or chatgpt.com) can only be approved by name, whatever its path.

Approving an assistant approves it for any account with that provider, including a personal one, because the console cannot tell a university plan from a personal account. If staff must use your institution's own plan, say so in your staff guidance.

Removing an assistant from the list ends every connection to it straight away, and each ended connection is a row in the audit log naming who removed it. Settings also shows how many attempts to connect an assistant that is not on the list were refused in the last seven days, and where they would have returned to.

Connecting

The server address is your console's address followed by /api/mcp, for example:

https://console.firstsix.com.au/api/mcp

In Claude, add it under Settings, Connectors, Add custom connector. Claude then opens the console's sign-in, you sign in with your university account as usual, and a consent screen shows the assistant's name, what it is asking for, and where it will send you back to. The name is whatever the assistant calls itself, so the return address is what decides: the screen offers Connect only when that address is one your institution has approved, and says so. Anything else gets a page saying it is not approved, with nothing shared and no link back to the assistant. Only continue if you started the connection from your own assistant in the last few minutes: anyone with an account at the same provider can open this screen and send you the link.

Each person can hold at most ten live connections. An eleventh is refused on the consent screen, and nothing is disconnected to make room: disconnect any you no longer use, then connect again. Your account shows how many you hold.

Every new connection sends you an email. It leads with the assistant as your institution approved it, worked out from the address it returns to, then says where it returns to, the name the assistant gave itself (which whoever set it up chose, so it never comes first), what it can do, when it was made, and a link to Your account. If it was not you, disconnect it there and tell whoever manages integrations. The email deliberately carries no one-click disconnect link, because a link that changes a connection without signing in would itself be a new way in. Whether the email went is recorded in your institution's audit log; if it could not be sent, the connection is still listed in Your account.

To disconnect, open Your account in the console and choose Disconnect next to the assistant. Someone who manages integrations can also see and disconnect every assistant connected at your institution. Both lists show where each connection returns to, which is how a genuine assistant is told apart from something that only borrowed its name.

How it is secured

  • OAuth 2.1 with PKCE, the standard MCP clients expect, with the console as its own authorisation server. Login is your institution's existing single sign-on; First Six never sees a password.
  • Public clients only. Assistants register themselves (dynamic client registration) and receive no secret, because PKCE is the protection. A registration grants nothing, is limited in size, cannot be made from inside a web page, and is removed after a day if nobody used it.
  • The consent screen cannot be skipped. Only the console's own server can issue the sign-in code an assistant exchanges for access, and only after the consent screen's signed approval, for the person and session that approved it.
  • The approved list is enforced by the database, not just the screen. It is checked when the sign-in code is issued, when it is exchanged, at every renewal and on every request, so a connection to an assistant your institution has since removed stops working at once. A safeguard on the connections table itself refuses any new connection to an address that is not approved, or over the limit below, whatever code path tries to create it.
  • At most ten live connections per person, checked on the consent screen and, for certain, in the database when the code is exchanged, under a lock held for that person, so two connections made at the same moment cannot both slip past it. The limit bounds how many connections one person, or someone holding their stolen console session, can create; it does not change what each connection can read.
  • Tokens are stored only as hashes. An access token lasts an hour. A refresh token lasts 30 days, is replaced every time it is used, and a connection cannot be extended past 90 days without connecting again. Presenting an old refresh token, however old, ends the whole connection, because it means two parties hold it. The one allowance is a single retry within 30 seconds of a renewal, by the same assistant, which is what a dropped network response looks like; if that retry was really a copied token, the connection ends the next time the genuine assistant renews.
  • Tokens are bound to this server (the resource parameter), so one issued for the console cannot be replayed somewhere else.
  • A made-up credential is refused before anything is looked up. Every sign-in code, access token and renewal token carries a signature only the console can produce, so a guessed or invented one costs nothing to turn away, and an access token carries its own expiry, so an old one is refused the same way. Limits on using a connection are kept per connection and per credential, not per network address, because everyone using a hosted assistant reaches the console from that provider's addresses, and a per-address limit would let one person's traffic lock out staff at every other institution. Claude and ChatGPT, registering at their own callback addresses, get their existing registration back each time they connect, at no cost. Any other registration (another hosted assistant, or an app on your own computer) is limited to 60 an hour per network address, so another hosted service's users share that limit. Approving a connection on the consent screen is limited per address too.
  • Text from university web pages is fenced. When the assistant reads your institution's connected sources, the page text arrives between two markers carrying a code made fresh for that response, with a note that it is material, not instructions, so a page cannot pretend the material has ended. Your institution's own names for its help categories, programs, courses and campuses arrive the same way, with only the keys a tool needs outside the markers, so a name written to read like an instruction is treated as a name.
  • Log out everywhere ends connections too, including one whose sign-in was approved before the log-out and exchanged after it.
  • "Off" stays off. When a permission is removed or a switch is turned off, each affected connection is ended the next time it is used, and does not come back if the permission or switch is restored.
  • Rate limits: 120 requests a minute per connection and 240 per person across all of their connections, plus limits on registration and the token endpoint (see Authentication, rate limits, and errors). If the shared store that keeps those counts is unavailable, the console refuses requests that make or use a connection rather than run them with weaker limits. Disconnecting is never refused this way.

Who processes what

Your institution decides whether its staff may connect an assistant and which assistants they may connect; each staff member then chooses among those. The list controls the provider, not the account: approving Claude approves every Claude account, a personal one included, so an institution that wants staff on its own plan should say so in its staff guidance.

The assistant's provider (for example Anthropic for Claude, or OpenAI for ChatGPT) processes what the assistant reads, under whatever terms apply between your institution and that provider, at your institution's direction. In that role it is not a First Six subprocessor: First Six does not choose the assistant or engage its provider for this connection, even where the same company supplies First Six's own writing assistant. Clause 3.7 of the data processing agreement sets this out. With content access, what it reads is your institution's authored content. With insights access it also reads exact aggregate figures. They never name a student, but a small group's figures can relate to an identifiable student, which is why the institution's approval and the staff member's permission both apply.

Common questions

Can the assistant publish to students by mistake?

No. There is no publishing tool on this server at all. Drafts reach students only when a person publishes them in the console.

Can it see a particular student's check-ins?

Not as a record: it reads only group figures and never a student's check-ins, name or words. The figures are exact, so in a very small group they can reflect how one student answered.

Which assistants can our staff connect?

Only the ones someone who manages integrations has approved in Settings, Integrations, each recognised by the exact address it returns to. With nothing approved, nobody can connect anything, and neither switch can be turned on.

What if someone connects an assistant in my name?

You get an email for every new connection (if it cannot be sent, the audit log says so, and Your account still lists the connection). If it was not you, open Your account, disconnect it, and tell whoever manages integrations: someone may have used your sign-in. No one can hold more than ten live connections, and every connection is in the audit log.

What does a demo account do?

A demonstration session cannot connect an assistant. Sign in with your own account.

How do we see what an assistant did?

In the console's audit log. Each connection, each call and each disconnect is a row, naming the person and the assistant, including disconnects the console makes by itself (for example when someone leaves, loses the permission, or a renewal token is used twice) and the ones that follow an assistant being removed from the approved list, which name who removed it.

Was this helpful?
Need more help?

The fastest answer is usually one question away.

Contact us
Edit this page on GitHub