Connect your AI

Connect your AI to copadre

copadre runs an MCP server — the open standard AI tools use to work with other products. Connect Claude to it and you can review conversations, draft and test agents, and publish changes without leaving the chat. Everything happens with your permission, and everything happens as you.

What the connection can do

Once connected, your AI tool can see and use your copadre account the way you can:

Review conversations. List them, read full transcripts, and ask the questions a list can’t answer — “what themes came up this month?”, “which questions couldn’t we answer?”
Check agent health. Out-of-knowledge questions, blocked visitors and broken hand-offs — the same events your dashboard shows.
Build and improve agents. Create an agent, save a new draft of its instructions, test the draft in a preview chat, and publish when it’s right — the whole loop, in conversation. Every AI-written draft records the request that produced it, so your version history shows what was asked alongside what was written.
Move agents as files. Export agents to a file and import them back — the same import and export the app offers, with the same safety rules.

It acts as you — never more than you. The connection carries your role in the app, exactly. Owners, developers and editors get the full set; viewers get the read-only view plus preview chat, just as they do in the app. There is no way to do something through your AI tool that your account couldn’t do in the app itself.

Connect Claude

In Claude, open Settings → Connectors → Add custom connector.
Paste the copadre server address:
```
https://bobby.copadre.com/mcp
```
Claude sends you to copadre to sign in — the same passkey or email link as always — and shows you exactly what the connection will be able to do. Approve it, and you’re returned to Claude, connected.

That’s the whole setup. There are no API keys to create or paste — the sign-in and approval are the setup. Custom connectors are available on Claude’s paid plans, in both the web app and the desktop app.

Claude asks before changing anything. Reading is quiet, but tools that write — saving a draft, publishing, importing — are flagged to Claude as changes, and it checks with you before running them.

Other AI tools

Claude is the tool we test with, but the server speaks standard MCP with standard OAuth sign-in, so any AI tool that supports remote MCP servers connects the same way: give it the same address, sign in, approve. If your tool asks for a “client id” or an API key, it’s using an older connection style we don’t support — look for its remote MCP server or custom connector option instead.

What the connection can never do

Worth knowing before you approve anything:

It never sees your keys. Your Claude API key and your embed keys are not readable through the connection, full stop.
It can’t manage your organisation. Team members, roles, billing, allowed domains and your organisation’s settings are app-only surfaces.
It can’t reach anyone else’s data. The connection is scoped to your organisation, and every request is checked against your account — live. Deactivate a user and their connection stops working on the next request.

Publishing is publishing. Asking your AI tool to publish an agent takes the draft live for your website visitors immediately — it is exactly the Publish button in the app. Draft and preview freely; treat publish with the same care you would in the app.

Preview chats route, just like the real thing. A preview is a real conversation with your agent, so if the agent you’re testing is a concierge — or its instructions hand off to other agents — the preview will route too, exactly as it would for a visitor. Every preview reply names the agent that actually answered and flags when it wasn’t the one you asked for, so your AI tool can tell a routed answer from the tested agent’s own knowledge. A sibling agent answering well is the cluster working — not proof the agent under test knew it.

If you want to test one agent’s knowledge boundary — what it does when it’s out of its depth — build it as a specialist and get it right first, then wire it into the cluster. Once an agent is part of a cluster you’re testing the cluster, which is the right thing to test later, but not while you’re still shaping a single agent’s instructions.

Staying in control

Every connection you approve is listed under Connected apps on your account page, with when it was last used and a Revoke button. Revoking cuts the connection off immediately. An approval also expires on its own after a year, so a connection you’ve forgotten about doesn’t live forever — you’ll simply be asked to approve again.

Try asking

Once connected, these are good first requests:

“What themes came up in my copadre conversations in the last 30 days?”
“Which questions did my agents fail to answer this week?”
“Draft a specialist for returns and refunds, let me try it in preview, and publish it when I say so.”
“Read my sales agent’s instructions and suggest where they’re vague.”
“Export my support cluster so we can work on it here.”

Where next

If you’re building agents through the connection, core concepts explains drafts, publishing and roles, and import and export covers the file format your AI tool can read and write. If you haven’t built anything yet, start with the starter cluster — then connect your AI and improve it from there.