MySendbox
Email inboxes for AI agents on your own domain: read, draft and reply, risky sends held for review. Remote server at https://mcp.mysendbox.com/mcp
Hosted MCP Server
npx add-mcp 'https://mcp.mysendbox.com/mcp'Installs into Claude Code, Codex, Cursor and more
Documentation
Bring your own agent.
Connect Claude, ChatGPT, Gemini CLI, GitHub Copilot, Cursor or your own code to a mailbox, with the same checks as our agent.
01
Before you start
- A MySendbox team with Agents turned on, and at least one mailbox.
- An owner on the team to approve the connection.
- For the agent to reply, not only read: open the mailbox's settings in the dashboard and set Who answers to “Your connected agent” or “Either”.
No account yet? Add the server anyway and choose “Create a workspace for your agent” when asked to sign in. It reads at once; replies start with the plan.
02
Add the server to your app
The server speaks Streamable HTTP and signs in with OAuth, so most apps only need the URL.
https://mcp.mysendbox.com/mcp
- Claude (web and desktop): add MySendbox from Claude's connector directory, or go to Settings → Connectors → Add custom connector and paste the URL.
- ChatGPT: add it as a custom connector (MCP server) with the same URL.
claude mcp add --transport http mysendbox https://mcp.mysendbox.com/mcp
gemini mcp add --scope user --transport http mysendbox https://mcp.mysendbox.com/mcp
{
"mcpServers": {
"mysendbox": { "httpUrl": "https://mcp.mysendbox.com/mcp" }
}
}
Then run /mcp auth mysendbox in Gemini CLI to sign in and approve the connection in your browser.
{
"servers": {
"mysendbox": { "type": "http", "url": "https://mcp.mysendbox.com/mcp" }
}
}
Choose Start above the server in the file, sign in when VS Code asks, then use it from Copilot Chat in agent mode.
{
"mcpServers": {
"mysendbox": { "url": "https://mcp.mysendbox.com/mcp" }
}
}
03
Approve the connection
The first time the agent connects, your app opens a page on app.mysendbox.com. Sign in as an owner, pick the mailboxes the agent may use, and choose its access:
- Read: read and search mail.
- Draft: read, tidy, and write drafts for a person on your team to send.
- Reply: read, tidy, and reply. Replies are checked, and wait for a person to review when they're risky.
04
Tools
read list_mailboxes list_messages search_messages
get_message get_thread get_attachment get_reply
account_status → what it may do in each mailbox, credits
wait_for → the reply it's waiting for, as it arrives
wait_for_code → the code or link from a sign-up email
list_threads → where each conversation stands
draft + mark_read move_messages draft_reply draft_email
+ set_thread → a note or a follow-up date on a conversation
reply + send_reply → sent, or waits for review
+ send_email → a new email: waits for a person to approve it,
unless the mailbox sends on its own
Signing up or logging in somewhere with the mailbox's address? Submit the form, then call wait_for_code: it waits up to 30 seconds for the email and returns its one-time code or verification link.
For mail that needs an answer, filter by category: "personal". The others are updates, newsletters and promotions.
Once it's connected, ask in plain words:
What came into support@ today that still needs an answer?
Draft replies to the unanswered order questions, for me to check.
Sign up for Acme with bookings@ and finish the email check.
Move this week's newsletters out of the inbox into Archive.
Ask three suppliers for a quote on 40 chairs, and tell me the cheapest when they answer.
Waiting for an answer? Right after send_email or send_reply, call wait_for with the conversation's thread_id: it returns the reply as soon as it arrives, up to 30 seconds a call, so your agent doesn't list the inbox again and again.
Every message comes with its facts: dates, amounts, tracking numbers, an invitation, and whether a person is asking something. Often that's enough to act on without reading the whole email.
Agents that run as servers can have a message.received webhook wake them instead; it carries the same facts.
Agents without a person to sign in
Servers and scripts can skip OAuth and send an API key as the bearer token. Create one in the dashboard under API keys, with the scopes the agent needs, or let the agent sign itself up and get a key with its own inbox:
Authorization: Bearer msb_live_…
messages:readto read,drafts:writeto draft,replies:sendto reply.messages:writeto tidy: mark read and move messages.
06
What stays safe
- Replies, and new emails a person approves. A reply goes to the person who wrote in, in the same thread, one per message. A new email goes from your own domain to at most 10 people, and waits for a person to approve it unless the mailbox is set to send on its own.
- Same checks, same approvals. Every reply is screened for leaks and for instructions planted in the email. Risky ones wait for a person, whichever agent wrote them.
- Outside mail stays outside. Messages reach the agent marked as untrusted, with their screening results. Quarantined mail is never shown.
- Every call on the record. Each tool call is logged with the agent, mailbox and outcome. Connections are limited to 120 requests a minute, on top of each mailbox's daily limits.
More on how we keep email wanted: Safety.
07
Disconnect
Owners see every connected agent under Connect agent, and can disconnect it at once.