# Agent Inbox: give your agent a way to reach you > Meet Agent Inbox: an inbox for AI agents on email, Telegram and Slack. Your agent can ask for approvals, wait for answers, and receive messages. Canonical: https://www.novu.co/blog/agent-inbox/ Markdown: https://www.novu.co/blog/agent-inbox.md Last updated: 2026-11-02T22:46:00.000Z Meet Agent Inbox: an inbox for AI agents on email, Telegram and Slack. Your agent can ask for approvals, wait for answers, and receive messages. Authors: Victor Yakubu Published: 2026-11-02T22:46:00.000Z Category: Agent Inbox At Novu, we've spent years building the infrastructure that gets messages to people on the channels they actually read. Novu Notify delivers product notifications across in-app Inbox, email, SMS, push and chat. Novu Connect brings AI agents into the channels people already use, like Slack, Microsoft Teams, WhatsApp, Telegram and email, so a product's agent can hold two-way conversations with its users. Both help a product reach its users and hear back. Today we're launching Agent Inbox, which does the same for the agent you run yourself: a coding agent, an assistant or a background job. Agent Inbox gives your agent its own identity on email, Telegram and Slack with more channels coming soon. It can ask you a question and wait for your answer, and people and services can write to it directly. We carry the messages. Your agent does the thinking. ## **Agents work outside the chat now** Agents used to live in a chat window. You typed, they answered, and you were right there when they needed something. That's not how most agents run anymore. They work in terminals while you're in meetings, on schedules overnight, and in CI jobs nobody is watching. That's when an agent most needs a person: - **Nobody is watching.** The run hits a question only you can answer, and it either guesses or stops. - **The action is hard to undo.** Deleting data, deploying to production, spending money, or sending something other people will read. - **A choice is costly to get wrong.** Two reasonable options, and picking the wrong one quietly means redoing the work. - **A long job is done.** Someone should know without checking a log. It runs the other way too. A service sends your agent a verification code. A teammate has a question about the report it filed. A vendor replies to the email it sent yesterday. Today, all of that lands in your inbox, and you forward it. So builders wire it up themselves: a mailbox for the agent, a Telegram bot, a Slack app, approval logic, somewhere to store the conversation, and code to match each reply to the question it answers. It's the same plumbing every time, and it isn't the part of the agent anyone wants to build. ## **Meet Agent Inbox** Every agent gets: - **A name and an email address,** like atlas@agentinbox.mx. Email works the moment the inbox exists. - **A contact card** with one link and a QR code, so people find every way to reach it in one place. - **Its own Telegram bot and its own Slack app,** when you connect them. Your agent never sends as you. Your agent can run in more than one place, like Claude Code on your laptop and Cursor at work. Each connects to the same inbox, so the agent keeps one address, one set of contacts and one history. ## **How it works** ### **Your agent asks, you answer** An agent can involve a person in four ways: - **ask:** a free-text question - **approve:** Approve and Deny buttons, for anything hard to undo - **choose:** up to 10 options, labels are customizable. - **tell:** an update that needs no reply The question arrives where you already read messages. The command waits until someone answers, then returns the answer with who gave it and when. ```bash $ npx @agentinbox/cli send approve "Ship v2.4.1 to production?" \ --from deploy-bot --timeout 15m Approved by dima at 14:02. $ echo $? 0 ``` Scripts branch on the exit code: 0 for approved, 10 for denied, 11 if the wait timed out, and 12 if the request expired. A timeout doesn't lose anything. The question stays open, and a late answer lands in the conversation. Send the question to several people and the first valid answer wins. In Claude, ChatGPT or Cursor, the same flow runs through MCP, and the agent keeps waiting until the answer arrives. ### People and services write first The inbox works in both directions. Anyone can email your agent's address, and once you connect Telegram or Slack, people can message it there too. A conversation stays unread until the agent replies, so nothing slips while it's busy. Before it starts something new, the agent checks what's waiting: ```bash $ npx @agentinbox/cli list --filter unread $ npx @agentinbox/cli reply "Yes, it's live." ``` Service email needs no forwarding. Sign up for Linear with your agent's address, and the verification code goes to the agent, not to you. In the dashboard, every conversation is sorted into New, Waiting and Resolved, and you can filter by channel. ### **You decide who can reach it** An agent with an address will hear from people you don't know, so the rules are built in: - **Contacts** are the people your agent can message. Invite someone with a link, valid for three days, and they connect the channel they prefer. - **Strangers** can write first if you allow it, and your agent can reply. It can't message a stranger first, and it treats a stranger's text as information, never as instructions. - **Contacts only** means messages from strangers never arrive. - **Pause** stops the agent from reading or sending. Messages wait until you resume. Your agent can't change any of these settings. Only you can, from the dashboard. ### **It works where your agent runs** Agent Inbox connects to the one you already have: - **The CLI,** for terminals, coding agents, scheduled jobs and CI. - **The MCP server** at https://mcp.agentinbox.md, for Claude, ChatGPT, Cursor and Grok Bot. - **Skills** for coding agents like Claude Code, Codex and Cursor. They teach the agent when to reach out and when not to, so it won't message you about something you could answer in the chat you're already in. ```bash npx skills add agentinboxhq/skills ``` ### **Messages can start your agent** A coding agent only reads its inbox while someone is running it. If a teammate writes at night, the message waits until you open a session the next morning. The listen command fixes that. Run it in a project folder, and when you, or a contact you choose, message the agent, your computer starts Claude Code in that project to handle the conversation and answer in it. ```bash npx @agentinbox/cli listen claude-code ``` - **You approve it first.** Turning on listening sends you an Allow or Deny message, and nothing starts until you allow it. - **You choose whose messages count.** Yours always do. A contact's do only after you turn on Listen too for them. A stranger's never do. - **Risky steps still ask.** Anything outside the project, like a command its settings don't allow, network access or messaging someone else, comes to your chat as Allow or Deny. - **It runs in the background** and starts again after a restart. listen status shows where it's listening, and listen stop turns it off. Claude Code is the first host listen supports. ## **Built for any agent that works with people** Agent Inbox fits any agent that needs to hear from people, or reach them. - **Sign-ups:** let your agent sign up for services and receive verification codes, receipts and invoices at its own address. - **Approvals:** get a yes or no before anything hard to undo, like a deploy, a refund or deleting data. - **Personal assistants:** let people message your agent directly on email, Telegram or Slack, and get a reply on the same channel. - **Team requests:** have your agent ask teammates, vendors or candidates for what it needs, and follow up until they answer. - **Reports:** send status updates, summaries and files when a job finishes, with replies kept in the same thread. - **Background jobs:** let a scheduled run or CI job check for new messages, pick up answers that arrived late, and carry on. ## **What's next** Agent Inbox starts with email, Telegram and Slack. Next on the list are WhatsApp, iMessage and voice, webhooks that wake your agent when a message arrives, custom email domains, and a REST API. We'll update the docs as each one ships. ## **Start today** Agent Inbox is free for 100 conversations a month, with no credit card required. Create an inbox from your terminal. You don't need an account to start: the inbox receives email straight away, and you claim it with an account when you want your agent to send. ```bash npx @agentinbox/cli setup ``` Or sign up at [agentinbox.md](), or add https://mcp.agentinbox.md to Claude, ChatGPT or Cursor. The [docs]() cover every command and tool. We'd like to hear what your agent does with its inbox.