The MCP server lets an AI assistant such as Claude, ChatGPT, Codex or Cursor work in your workspace: find contacts, write and submit templates, build and test chatbots, manage webhooks and triage the inbox. It uses the same tools as the in-panel Copilot. Each call runs as you, so your permissions, your plan and the audit log apply.
Where to find it
You connect from your AI client, not from the panel. You need:
- the MCP server address for your brand (ends in
/mcp);
- a panel account that is a manager or has the Connect AI tools permission.
The address is shown in the panel under Sidebar → Copilot → Connect an AI app, in Your server address, with a Copy address button. Always copy it from there; it is different for each brand. See Connect your AI app.
Before you start
- Your plan must include MCP access (
mcp_access). Without it every tool answers entitlement_required; only whoami works, and it says why.
- Some tool groups need their own plan features (for example webhooks or AI agents). Tools you cannot use are hidden.
- Sign in with OAuth when your client supports it. It needs no token copying.
Steps
Claude (claude.ai and Claude Desktop)
In Claude, open Customize → Connectors → Add custom connector (older Claude apps: Settings → Connectors)
Step-by-step for every AI app: Connect your AI app. On Team or Enterprise plans an owner adds it once, then members click Connect.
Enter a name and the MCP server address
Leave the advanced OAuth fields empty.
A sign-in page opens
Sign in with your panel account.
On the consent page, untick anything you do not want the assistant to do, then click Allow
The page lists each permission the assistant asks for, each with a checkbox.
In a chat, turn the connector on and ask "Who am I?" to run whoami
Claude Code
- With the plugin: add the plugin's marketplace, install the plugin, run
/mcp and choose Authenticate. The plugin adds skills and slash commands for chatbots, templates, broadcasts, webhooks, inbox triage and diagnosis.
- Server only:
claude mcp add --transport http <name> <server address>, then /mcp → Authenticate.
ChatGPT, Codex and other clients
- ChatGPT: turn on developer mode in Settings → Apps & Connectors → Advanced settings, create a connector with the server address and OAuth, then sign in and allow.
- Codex: add the server under
mcp_servers in ~/.codex/config.toml, then run codex mcp login <name>.
- Anything else (Cursor, VS Code, MCP Inspector): any client that speaks MCP Streamable HTTP works. Clients without OAuth send
Authorization: Bearer <token> on every request.
Permissions you can grant
| Permission | Lets the assistant |
|---|
See your workspace (mcp:read) | Read your data. Always on |
Edit contacts (crm:write) | Create and change contacts, tags, groups, fields |
Manage messages and templates (messaging:write) | Templates, quick messages, sends |
Build chatbots (bots:write) | Create and edit chatbots |
Build AI agents (agents:write) | Create and edit AI agents |
Manage webhooks (webhooks:write) | Endpoints and subscriptions |
Manage mini apps (apps:write) | Install, configure, remove apps |
Work in the inbox (inbox:write) | Assign, note, tag, close |
Connect apps for chatbots (creds:write) | Save and test app keys (never read them back) |
Build webhook automations (scenarios:write) | Automations that start from a webhook |
Manage your knowledge base (kb:write) | Knowledge base sources and settings |
Manage your AI provider keys (aikeys:write) | Set, test, remove AI keys. Managers only |
Edit business hours (hours:write) | Working-hours calendars and holidays |
Catalog and order permissions (ecom:write, orders:write) also exist for workspaces with ecommerce. If the assistant later needs a permission you did not grant, it asks you to sign in again.
Approvals
- Changes that are destructive or visible to customers need the assistant to confirm them.
- In auto mode (the default) a confirmed change runs at once and is recorded.
- In ask mode the change waits for you. You get an approval link and approve it in the panel; the assistant then retries.
- Some changes always need a person, whatever the mode: deletes, uninstalling an app, rotating a webhook secret or credential, removing an AI key, stopping or cancelling a broadcast.
- A manager sets the workspace mode in the Copilot settings. Any staff member can choose "ask me first" for their own calls; nobody can make it less strict.
- A real broadcast send always needs your approval with your two-factor code in the panel. Never type a 2FA code into a chat.
Limits and plans
- MCP access depends on your plan (
mcp_access). Check Billing → Current Plan.
- The MCP server shares rate limits with the REST API for your workspace.
- OAuth access tokens last 10 minutes and the client renews them itself.
Good to know
- New chatbots made by an assistant are created switched off. Turning one on is a separate, confirmed step.
- Secrets are never returned. A webhook signing secret is shown once.
- Data from your workspace (messages, notes) is treated as data, never as instructions to the assistant.
Do it with Copilot
The same tools run inside the panel's Copilot. Try: "What can you do in this workspace?", "List changes waiting for my approval" (approvals_list_pending).
Troubleshooting
| What you see | Why | What to do |
|---|
entitlement_required | MCP access, or that tool's feature, is not in your plan | Run whoami to see which; check Billing |
insufficient_scope | You did not grant that permission | Reconnect and tick it on the consent page |
confirm_required with a link | Workspace is in ask mode, or the change always needs a person | Open the link, approve in the panel, ask the assistant to retry |
| Sign-in loops or fails | Wrong server address, or your account lacks Connect AI tools | Check the address; ask a manager for the permission |