AI assistants (MCP)

Buzzmark runs a mailbox MCP server at https://buzzmarkapp.com/api/mcp/mail. Connect an assistant to one hive and it can read the mail that hive receives, reply to it, and send new email - only ever as the address you choose, and only within the policy you set.

Connect in a minute

claude.ai and Claude Desktop

1. Open Settings → Connectors → Add custom connector.
2. Paste the URL https://buzzmarkapp.com/api/mcp/mail and leave the OAuth fields empty.
3. Sign in to Buzzmark when asked, then choose the hive, the From address and what the assistant may do.

Claude Code, Cursor and other clients

Open the hive in the dashboard, go to AI access and create a key. The page gives you the exact command or config to paste. For example:

claude mcp add buzzmark-mail --transport http https://buzzmarkapp.com/api/mcp/mail \
  --header "Authorization: Bearer bzm_mbx_..."

Any MCP client works: use the URL above with an Authorization: Bearer header.

Choose what it can do

Every connection has its own settings, and you can revoke it at any time - it stops working on the very next request.

Sending option	What happens
Ask first (default)	Every email waits in the hive’s AI access page until a person approves it. You get an email when something is waiting.
Replies only	The assistant can reply to emails the hive received, and nothing else. Good for support inboxes.
Listed recipients only	Sends straight away, but only to the addresses or whole domains you list.
Anyone	Sends straight away to any address.
No sending	Read-only.

You also choose whether it can read the hive’s incoming mail, its daily limit (up to 1,000), and - for keys - when it expires.

Tools

An assistant only sees the tools its settings allow.

Tool	What it does
get_mailbox	The From address, where the hive receives mail, the sending policy and remaining limits
list_inbox	Received emails, newest first, with a short preview; filter by sender, subject or date
read_email	One received email in full, as plain text, with attachment names
reply_to_email	Reply to a received email - threaded, to its Reply-To or sender, with a “Re:” subject
send_email	A new email to up to 10 recipients (To and Cc)
list_sent	What it has sent, with delivery status, plus anything waiting for approval or rejected

How it stays safe

• One hive, nothing else. A connection can’t see other hives, change settings, read your account data or use the account-management API - and account tokens don’t work on the mailbox server.
• A fixed From address on one of your verified domains. The assistant can’t change it.
• No BCC, no attachments, no custom headers, at most 10 recipients per email and 10 emails a minute, plus your daily limit.
• Every account rule still applies: suppressions, verified-domain checks, rate limits and plan allowances.
• Defences against manipulation. Incoming email is labelled as untrusted for the assistant and stripped of hidden characters, and it won’t reply to auto-replies, bounces or no-reply addresses - so it can’t be talked into a mail loop.
• A full trail. Everything an assistant sends is tagged ai-assistant and shows which connection sent it; approvals record who approved them.
• Keys are stored hashed and shown once. Connector logins use OAuth 2.1 with PKCE, and the grant covers only the hive you picked.

An assistant that reads your inbox can still be misled by a cleverly written email. If it handles anything sensitive, keep Ask first or Replies only on, and read emails before approving them.

Managing accounts too

Buzzmark also has an account-management MCP server at https://buzzmarkapp.com/api/mcp for creating hives and querying logs and stats, authenticated with your account API token. It can’t send or read mail; the mailbox server above can’t manage the account. Keep them as separate connections.