Connect your agent to Claude

What you get

Your agent already talks to customers all day. This connects that pile of real conversations to your own Claude, so you can ask about it the way you would ask a colleague — and Claude can go and fix what it finds, without you opening a settings page.

It works over MCP (Model Context Protocol), the standard way an AI assistant plugs into an outside system. In practice: you paste one address into Claude, sign in, and a set of tools appears that Claude can use on your behalf. Nothing is installed, and nothing runs on your computer.

There are three connectors, and they are separate on purpose:

• Agents — your conversations, contacts, channels and the agent's own knowledge. This is the one this guide sets up.
• Sales — groups, journeys and audiences in the Sales product.
• Tasks — your task board.

Before you start

Three things have to be true. The third one is the one that trips people up, so check it even if you are sure.

• An Assistant Labs account with at least one agent. If you have not built one yet, start with the 3-week playbook instead — there is nothing to ask Claude about until the agent has had some conversations.
• A Claude account. Any paid plan on claude.ai can add a custom connector. Claude Code works on the same account.
• An organization, with your agent in it. A connection is granted to one organization's agents, so an agent that belongs to no organization cannot be shared with anything.

If you do not have an organization yet

Open Account → Organization. If the page says you have none, click Create Organization, name it after your business, and you become its admin.

Then add the agent: creating the organization does not move your existing agents into it. On the same page, under Assistants, use Add assistant… and pick the agent you want to connect. If that list is empty, the connect screen will keep telling you there is nothing to share.

Only if you plan to use an API key

Signing in needs nothing more than the above. The API-key path is different: keys belong to the developer API, which is off until our team switches it on. Open Account → Developers, click Request developer access, say in a line what you want to build, and you will get an email — usually within one business day. Only an organization admin can request it, and only an admin can create keys afterwards.

Find the connect screen

Everything you need — the address, the key button, and the list of what is already connected — is on one page inside the agent you want to share.

The same panel appears a second time under Account → Developers → MCP, for the whole organization rather than one agent. Either one works; the per-agent page is easier because it already knows which agent you mean.

Now make the one choice the rest of this guide follows. Everything below switches with it.

Connect claude.ai and the desktop app

Same screen in both — Claude Desktop's settings are the web settings. Do it once on a computer and it works in the Claude app on your phone too, on the same account.

Claude’s own menus are in English even if you use Assistant Labs in Hebrew, so the button names below are written the way they appear on screen.

Connect Claude Code

Claude Code takes the same connector from the terminal — useful when you want the agent's data next to a codebase, or want to script something against it.

What to ask it

The connection is not a chat window onto your database — it is a set of tools Claude picks from. You describe what you want in ordinary language and it works out which ones to call.

Reading what actually happened

• “What did customers ask me this week?”
• “Which conversations got no reply?”
• “What time of day do most messages come in? Should I be staffing differently?”
• “Read the last 50 conversations and give me the five questions the agent answered badly.”
• “Has anyone complained about delivery times this month?”

Fixing the agent while you talk to it

This is the part people underestimate. Claude does not just report — with the edit permission on, it can change what the agent knows and the change applies to real conversations from that moment.

• “Customers keep asking whether we ship to Eilat. Add it to the FAQ and show me exactly what you changed.”
• “The agent is too formal. Loosen the tone but keep it polite.”
• “Re-read our website and update the business details from it.”
• “Simulate a customer asking about returns and show me what the agent says.”
• “File a training suggestion for that bad answer instead of changing it live — I want to review it first.”

Training suggestions are review-first: they land in the Training tab for you to approve. Nothing goes live on its own.

Contacts, channels and templates

• “How many new contacts came in this month, and where from?”
• “Which channels are connected right now?”
• “Draft a WhatsApp template apologising for a delayed order, then create it for me to submit.”

On very large questions: answers are capped so a single question cannot pull an unbounded amount of transcript — roughly 300 conversations per analysis and 200 messages per conversation. When a result is cut short it says so and asks you to narrow the dates, rather than quietly answering from half the data.

Sales and Tasks

Same procedure, different address. Add them as two more custom connectors, each with its own approval.

Sales

Groups, journeys and audiences: “which journeys are running and how are they doing?”, “build me a win-back journey for customers who have not ordered in 90 days”, “who is in the hot segment right now?”. Planning and drafting are granted by default; taking a journey live is not — that sends real marketing messages and cannot be recalled, so it is a separate tick.

Tasks

Your board: “what is stuck?”, “open a task to call the supplier back, due Thursday”, “summarise everything assigned to me this week”. Both reading and writing are granted by default here — nothing on a task board reaches a customer.

Disconnecting, and when it breaks

Access you cannot take back is not access you should have given. Both paths have a one-click undo, in the same panel you connected from.

• If you signed in — the connect page lists Connected apps: who is connected, what they may do, and when they last used it. Disconnect takes effect on that app's very next request.
• If you used a key — revoke it from Account → Developers → API keys. Every connection made with that key dies with it.

• “No agent to connect” on the approval screen — your account reaches no agent that belongs to an organization. Go back to Before you start and add the agent to your organization.
• Claude says it is not authorised — the key was revoked, or the sign-in was disconnected. Reconnect; nothing is lost.
• Claude refuses one particular request — that permission was not ticked at approval time. Disconnect and reconnect, ticking it.
• The connector was added but the chat ignores it — it is off in that conversation. Click the plus next to the message box, choose Connectors, and switch it on. It is per-chat, not global.
• You pasted the address and nothing happened — check you copied the whole thing, including /mcp at the end.

Next, the same job on the other side of the fence: connecting to ChatGPT and Codex.