Skip to content
All help guides

Connected tools

Use your clients in Claude

This guide shows you how to connect Claude or Claude Code to AgentHub, pick the clients it can read, and stop it when you want to. Client records stay read-only.

7 minWorkspace owners, admins, and operators the owner allowsReviewed November 7, 2026

Before you start

  • You are a workspace owner or admin, or an operator whose owner turned on Let operators export client context and connect AI apps in Settings, Team.
  • You have permission to share these clients' information with your AI app.
  • For Claude: on a Free plan you can add one custom connector. On a Team or Enterprise plan, an authorized administrator of your Claude organization adds it. After that, you click Connect.

Open AI apps and copy the address

Open Settings, then Tools and apps, then the AI apps tab. Click Connect Claude. Under Setup details, click Copy. The address is https://mcp.agenthub.ai/mcp.

Open AI apps

Add AgentHub in Claude

In Claude, open Customize, then Connectors. Click + Add, then Add custom connector.

Enter the name AgentHub. Paste the address into MCP server URL, then click Continue.

Claude detects the settings. Keep Sign in now and Use Claude's published identity. Register automatically does not work with AgentHub.

Click Add, then Connect.

Important: On a Team or Enterprise plan, an authorized administrator adds it under Organization settings, then Connectors. See Anthropic's steps. Everyone else finds AgentHub under Customize, then Connectors, and clicks Connect.

Or add AgentHub to Claude Code

If you sign in to Claude Code with a Claude account and you already added AgentHub in Claude, Claude Code may already have it. Run /mcp in Claude Code to check. It then reads through your Claude connection, so you can skip the commands below.

Otherwise, run this command. It adds AgentHub to all your projects: claude mcp add --scope user --transport http agenthub https://mcp.agenthub.ai/mcp

Then run this command. Your browser opens so you can sign in: claude mcp login agenthub

Pick your clients

AgentHub opens Connect Claude to AgentHub, or Connect Claude Code to AgentHub. If you are not signed in, sign in first. The connection uses the AgentHub account your browser is signed in to, so check that account if you have more than one.

Tick the clients the app may read, or click Select all. Read the notice under the list, then click Connect. To stop here, click Cancel. Nothing is shared.

You see this page each time you connect.

Use it in Claude

In a chat, click +, then Connectors, and turn on AgentHub. Then ask about a client. For example: “What is open for Lantern Bay?”

Claude can ask before each read. That depends on your Claude mode and on the connector's Tool permissions. On a Team or Enterprise plan, your organization can set those. Claude Code needs your permission for a tool in its default mode. AgentHub does not control these settings.

Change clients

On AI apps, click Change clients, change the ticks and click Save. A client you add works right away. Taking a client away disconnects the app until you reconnect it.

Stop the connection

On AI apps, click Disconnect, then Disconnect again to confirm. That connection's next read is refused. Removing AgentHub inside Claude or Claude Code does not end it, so use Disconnect in AgentHub.

Claude and Claude Code are separate connections. Claude Code can also read through your Claude connection, so disconnect both to stop both.

A workspace owner sees everyone's connections and can click End connection on any of them.

What a connected app can do

A workspace member can connect Claude or Claude Code to AgentHub, so the app can read the clients that member picks. Client records stay read-only. The app can read recorded approvals, checks and published updates. It cannot approve, check, publish or send anything.

Who can connect

Owners and admins can connect. Operators can if the workspace owner allows it. Viewers cannot. Each person signs in to AgentHub, picks clients and clicks Connect. Nothing connects before that. Each person has one connection for each app. The workspace owner can see everyone’s connections and end any of them.

What the app can read

For each client the member picks:

  • The client’s name, and whether it is current or historical.
  • Open requests: title, scope, status, due date, approval state and date, approved scope, and any recorded check. AgentHub labels each request as coming from a member, a client or AI. The label records how the text entered AgentHub. It does not prove who wrote each word.
  • Published client updates: title, period, summary, the consultant’s note, the changes included and the items reported.
  • The open promises AgentHub saved from that member’s own conversations about the client, with their due dates and the date of the note behind them.

What that text can carry

This text can carry more than you might expect:

  • A client’s own words go out once a person has confirmed them as a request.
  • A request AgentHub helped prepare can include details from the client’s workflows.
  • A published update keeps the words of the requests it covers, even after a request closes.
  • A saved promise can repeat words from a conversation.

What the app cannot read directly

  • Closed requests, as records of their own.
  • Drafts, and updates that were never published or were revoked.
  • Contact fields.
  • Conversations with AgentHub.
  • Internal notes, uploaded files, evidence, workflow records and setup summaries.
  • Content imported through AgentHub’s Slack or Google connections or from the marketplace, and anything AgentHub can trace to that content.

Anyone can repeat these details in a request, an update or a note. Those words then go out like any other text.

Text AgentHub holds back

AgentHub shares a record’s text only when its source rules allow it. AgentHub started tracking where text comes from around October 31, 2026.

  • The text of a request from before tracking started stays in AgentHub. So does the text of a request that came from Slack, Google or the marketplace. Editing such a request does not release it.
  • AI-written request text and AI-written update summaries go out only from a workspace created after tracking started. That workspace must also never have connected Slack or Google or used marketplace content. Connecting Slack or Google sets this limit for good, before anything is imported. Deleting the connection does not lift it.
  • A client update’s text stays in AgentHub if it covers a held-back request. Outside such a workspace, it also stays if it has an AI summary or an item AgentHub cannot trace.
  • Saved promises, and update items drawn from a person’s notes, can go out from any workspace when AgentHub can trace them to that person’s own words.

When text is held back, the app still sees that the record exists, with its status and dates.

AgentHub cannot tell where typed words came from. Words someone pastes into a new request, from Slack, Google or anywhere else, count as that person’s own text, and the app can read them.

The filter, and its limits

Before text leaves AgentHub, a limited filter cleans it. The filter removes hidden characters, HTML comments and images, and links in formats it recognizes. It masks email addresses and phone numbers in formats it recognizes. It holds back a whole field if the field looks like a credential, a US Social Security number or a card number, or if AgentHub cannot clean it safely.

This filter is a fixed set of patterns. It is not complete, and it will miss things:

  • Names, street addresses and other personal details in ordinary sentences go out as written.
  • Health information has no detector.
  • Government ID numbers have no detector, apart from the US Social Security format. A nine-digit number with no separators goes out.
  • A credential in a format AgentHub does not know can go out.
  • A card number in an unusual format can go out, for example with a comma and a space between groups.
  • An email address in an unusual format can go out, for example dana [at] example [dot] com.
  • A phone number in an unusual format can go out at any length, for example with underscores. So can a phone number with fewer than 10 digits.
  • A bare web address or file name stays in the text, such as example.com or orders.csv.
  • Text hidden by formatting can stay. The filter does not find every hidden instruction.

Keep health data, government ID numbers, card numbers and credentials out of any client you connect.

Where the text goes

The connected app and the AI model service it is set up to use receive the text. With a Claude sign-in, that service is Anthropic. The account can belong to your organization. An organization can also set up Claude Code to use another cloud provider. The terms and settings of that account or organization apply, including training and how long the text is kept. AgentHub cannot see those settings, and does not control what the app does with the text.

AgentHub’s tools ask your AI app only for a client and a record. They do not ask for your chats, memory or files.

Words in a client record can hold instructions, and an AI app can follow them by mistake. AgentHub’s tools cannot change a client record. The app may have other tools connected, and AgentHub does not control those. Check what an AI app proposes before you act on it.

What AgentHub keeps

  • For each connection: the person, the app, the clients picked, its status and its dates.
  • A hash of each sign-in token, never the token itself. Disconnect deletes the hashes at once. Some used-token hashes can stay for the sign-in’s 180-day lifetime, so AgentHub can catch a second use. A daily job deletes them after that lifetime and one more day.
  • One access record for each read: the person, the connection, the client, the tool, the time, and how many records were read or held back. It holds no record text.

A daily job deletes access records once they are more than 90 days old, and an ended connection’s details 90 days after it ended. Deleting a closed workspace deletes them too. These reads make no AI call inside AgentHub and do not use your AgentHub AI allowance.

What Disconnect stops

Disconnect is in Settings, on the AI apps page. It ends that one connection. AgentHub deletes the connection’s sign-in tokens at once, and that connection’s next read is refused. A read that has already started can finish.

What Disconnect does not do

  • It does not recall anything. Whatever the app has already read stays where the app or the person put it: chats, memory, projects or anywhere else. AgentHub cannot reach it. Use the app’s own controls to remove it.
  • It does not end your other connections. Claude and Claude Code are separate connections. If you sign in to Claude Code with a Claude account, Claude Code can also use the AgentHub connector you added in Claude, through your Claude connection. To stop both, disconnect both.
  • It does not end anyone else’s connection.
  • It does not delete the access records or connection details above. They stay until the daily job removes them.
  • It does not remove AgentHub from the app. The app can still list AgentHub, and that connection cannot read.
  • It does not stop the same person from connecting again.

Removing or disconnecting AgentHub inside Claude or Claude Code does not end the connection in AgentHub. Those apps have no way to tell AgentHub. Use Disconnect in AgentHub to end it.

What else ends access

  • Taking a client away with Change clients ends the app’s access until the person reconnects. Adding a client takes effect with no new sign-in.
  • Connecting the same app again replaces the person’s earlier connection.
  • AgentHub ends a person’s connections when they leave the workspace or lose a role that can connect.
  • AgentHub ends every Operator’s connection when the owner turns off Let operators export client context and connect AI apps.
  • AgentHub ends every connection when the workspace is archived or starts closing.
  • A connection with no use for 30 days needs a new sign-in. Every connection needs a new sign-in within 180 days.

Verify it worked

  • AI apps shows the app, Connected, the last time it was used, and your clients.
  • Ask the app to list your AgentHub clients. It names only the clients you ticked.
  • Ask about a client you did not tick. The app gets “Client not found or access denied.”

Common issues

I disconnected in AgentHub, and Claude still shows AgentHub as connected

Claude keeps its check mark. That connection cannot read. To use it again, click Disconnect, then Connect, in Claude.

I disconnected Claude Code, and it can still read

Claude Code may be using your Claude connection. On AI apps, disconnect Claude too.

The app says "This connection is not active. Reconnect AgentHub."

The connection ended, or needs a new sign-in. In Claude, click Disconnect, then Connect. In Claude Code, run claude mcp login agenthub.

AI apps says "Reconnect from the app to use this again"

Reconnect from the app, as in the line above. Your earlier ticks show again. Check them before you click Connect.

The app says to turn on AI text sharing, or a record shows no text

AgentHub has no AI text sharing setting, and no setting releases withheld text. When AgentHub withholds text, it tells the app why with the record. See Text AgentHub holds back and The filter, and its limits. Editing does not clear a source rule.

The text says [email removed], [phone removed] or [link removed]

The filter masked it. The full text is still in AgentHub.

Claude says a connector with this URL already exists

You or your organization may already have added it. Find AgentHub in the connectors list and click Connect.

Claude Code stopped working on another computer

You have one connection for each app. A new sign-in replaces the last one. Run claude mcp login agenthub on the computer you want to use.

Claude Code cannot find agenthub

You may have added it without --scope user, so it works only in that project. Run the command in the Claude Code step again.

The page says "You cannot connect this workspace"

Ask the workspace owner. Viewers cannot connect. Operators need the owner's permission.

The page says "This sign-in has expired"

Go back to the app and connect again.

The app says "Too many calls"

Wait for the seconds it names, then try again.

Related guides