AI Connectors

Connect Claude or ChatGPT to ShareAndGo

Bring your own AI assistant into ShareAndGo via the Model Context Protocol. Ask questions, search across every room you have access to, and (with write scope) let your AI create rooms and upload documents on your behalf. All without copying a single file.

Two consent gates, by design

AI access in ShareAndGo is controlled at both the connector level (you decide what your AI can do globally) and the room level (each owner decides what AI can do in their specific room). Your AI can only act when both gates are open.

Token scopes (yours)

When you connect your AI, the consent screen shows two checkboxes:

  • Read access: list rooms, read documents, search, ask questions
  • Write access: create rooms, upload, move, delete documents

Write defaults to OFF. It's a deliberate opt-in.

Per-room toggles (theirs)

Each room has separate Read access and Write access toggles, controlled by the room owner from the Insights tab. Recipients can request read access; owners must enable write themselves.

Workspace defaults live at Settings → Workspace.

How to connect

From start to working connector in under a minute.

1

Generate your MCP URL

Open Settings → Integrations in ShareAndGo and scroll to the AI Assistants section. Click Connect AI, give the connector a name, and copy the MCP server URL.
2

Add the connector to your AI

Which clients need a key, and which do not. This trips people up, so it is worth knowing before you start. claude.ai and ChatGPT never take a key. You give them the server URL, they send you to ShareAndGo to sign in, and you approve the connection on a consent screen. Their “Add custom connector” dialogs have no field for a key at all, so if you are looking for one, nothing is missing. Claude Desktop and other MCP clients do use a pasted key, because they have no browser to sign you in with. ShareAndGo shows you a key either way; it is only needed for that second group.
claude.ai (paid plans): Settings → Connectors → Add custom connector. The dialog asks for two things only, a Name and an MCP server URL. Paste the URL and press Continue; Claude handles the sign-in and consent for you. There is no key to copy across.
ChatGPT (Business, Enterprise or Edu): Settings → Plugins → Add an MCP connector → paste the URL, then sign in and approve when ChatGPT sends you to ShareAndGo. There is no key to copy across. ⚠️ Not available on ChatGPT Plus — the option is simply not offered, so if you cannot find it on a Plus account it is a plan limit at OpenAI's end rather than a menu you have missed. Verified on our own accounts 28 September 2026; OpenAI move this, so check the date.
Claude Desktop (this one does use the key): Settings → Developer → Edit Config, and paste the JSON snippet shown after you click Connect AI in ShareAndGo. The snippet already has your key in it, so you do not need to copy it separately. Restart Claude Desktop afterwards.
3

Approve scopes on the consent screen

Your browser will land on a ShareAndGo consent page showing your email and two checkboxes: Read (default on) and Write (default off). Tick what you want and click Authorize.
4

Enable AI on a specific room

Open any data room → Insights tab → flip the AI Assistant Access toggle (Read or Write). Your AI client can now see and act on that room.
5

Try it

In Claude or ChatGPT, ask "What ShareAndGo data rooms can I see?" The connector will call list_data_rooms and show the rooms you've granted access to.

What your AI can do

Read tools work as soon as a room owner enables AI read access. Write tools additionally require AI write access on the room AND the write scope on your token.

list_data_roomsread
List every room with AI access enabled, that the connector's owner can see.
list_documentsread
List documents in a specific room with filenames, AI tags, summaries, and IDs.
search_documentsread
Free-text search across one room or the entire AI-enabled tenancy.
read_documentread
Pull the full extracted text of a specific document.
ask_roomread
Ask a natural-language question. ShareAndGo's AI synthesises an answer from the room's documents.
create_data_roomwrite
Create a new room in the connected user's organisation. Plan limits apply.
upload_documentwrite
Upload a base64-encoded document into a room (max 10MB per file).
create_folderwrite
Create a folder, optionally nested under a parent folder.
move_documentwrite
Move a document to another folder (or to room root).
delete_documentwrite
Soft-delete a document. Recoverable from the Recycle Bin by an admin.

What works in which client

The MCP transport itself is universal, but each AI client's sandbox has different limits. Here's what's reliable today.

ClientRead / search / Q&ASmall writes (rooms, folders, metadata)File uploads
claude.ai (web)✓✓⚠ Slow / unreliable
Claude Desktop✓✓✓
ChatGPT Desktop✓✓✓ when MCP ships generally
Custom MCP clients✓✓✓

Why claude.ai web is slow on uploads: Anthropic's web sandbox blocks external HTTPS hosts, so the only file-upload path is inline base64 in the tool-call argument. Claude has to materialise the file character-by-character. Fine for a few KB, painful past ~75KB, unreliable past ~150KB. We'll lift the limit when Anthropic improves the sandbox transport. For now, the ShareAndGo web UI is the right place to upload files; let your AI handle everything else.

Example prompts to try

Real prompts that work cleanly in claude.ai web today.

"What ShareAndGo data rooms can I see?"
"Summarise all documents in the BetaCorp ATO Audit room."
"Find every document that mentions 'overseas contractor' across my tax audit rooms, and quote one source clause from each."
"Read the BAS for Q2 2026 and tell me the GST payable."
"Create a new room called 'EOFY 2026 Acme Ltd'."
"Compare the indemnity clauses in the SPA and the Master Services Agreement in the Acme room."

For bulk uploads, document moves at scale, and similar byte-heavy operations, use Claude Desktop, ChatGPT Desktop, or a custom MCP client. See the table above.

Every action is audited

Every tool call your AI makes is recorded in two places: ShareAndGo's global ai_query_logs collection, and the affected room's tamper-evident audit log. Look for actions starting with ai_mcp. in the room's Audit Log tab.

ai_mcp.list_data_rooms · dana@example.com via Claude · 2026-05-08 16:42
ai_mcp.search_documents · dana@example.com via Claude · 2026-05-08 16:43
ai_mcp.read_document · dana@example.com via Claude · 2026-05-08 16:43
ai_mcp.upload_document · dana@example.com via Claude · 2026-05-08 16:45

AI-issued tokens can be revoked instantly from Settings → Integrations → AI Assistants. Revocation takes effect on the next request, with no propagation delay.

Troubleshooting

I generated a key but there is nowhere in ChatGPT or Claude to paste it

Nothing is wrong and you have not missed a screen. The “Add custom connector” dialogs in claude.ai and ChatGPT ask only for a name and a server URL. Those clients sign you in through ShareAndGo instead of taking a key, so the key you were shown is not used there at all. Paste just the server URL and press Continue. The key exists for Claude Desktop and other MCP clients that have no browser to sign you in with.

I cannot find "Add an MCP connector" anywhere in ChatGPT

Custom MCP connectors require a ChatGPT Business, Enterprise or Edu workspace. They are not offered on ChatGPT Plus, so on a personal subscription the option genuinely is not there rather than hidden somewhere. That is a plan limit at OpenAI’s end and not something we can enable. If you are on a personal subscription, Claude on a paid plan connects today with no admin approval and nothing in beta. Checked on our own accounts on 28 September 2026 — OpenAI have moved this before, so if it looks different, trust what you see.

My AI says the server is unreachable, or returns something that is not JSON

Check the address you pasted. The MCP server lives at /mcp/v1, and the URL shown in Settings → Integrations is always the right one to copy. A URL pointing at a page rather than the MCP endpoint will return a web page, which most clients report as the server being broken rather than as a wrong address.

My AI says "no rooms accessible" but I have rooms

This is the per-room consent gate working as designed. Open the room → Insights tab → flip the Read access toggle on. Your AI will see it on the next call.

My AI says "this connector does not have write scope" when creating a room or moving files

You only granted read scope when you authorized the connector. Revoke it from Settings → Integrations → AI Assistants and re-add it, ticking the Write checkbox on the consent screen.

Uploading a file via claude.ai is taking forever or fails

That's a known Anthropic-sandbox limitation, not a ShareAndGo issue. Claude.ai's web client has to base64-encode the file into the tool-call argument character-by-character. Files past ~75KB get slow, past ~150KB get unreliable. For uploads at any size, use the ShareAndGo web UI directly, or switch to Claude Desktop / ChatGPT Desktop / a custom MCP client where uploads work normally.

"AI write access not enabled for this room"

Your token has write scope but the room owner hasn't turned on the room-level write toggle. Open the room → Insights tab → enable Write access (admin only). Both gates must be open.

claude.ai shows "Sign in to ShareAndGo first"

Open shareandgo.com.au, sign in, then re-add the connector in claude.ai. Sessions are scoped to the frontend domain so the consent flow needs an active ShareAndGo session.

Privacy posture

ShareAndGo doesn't train any models on your data. The MCP server simply hands off document content to whichever AI you've connected (Claude, ChatGPT, etc.); that AI's own privacy policy applies to anything it sees. The connection is identical in privacy posture to downloading a document and pasting it into the AI yourself, except more secure: no laptop copy, no email forwarding, full audit trail. The room owner controls whether AI access is even possible.

Ready to bring your AI to ShareAndGo?

Set up takes about 60 seconds. Read access is on by default; write is opt-in. Revoke any time.

Connect your AI