We do not build inboxes for humans first
Human mail clients optimize unread badges. Agent inboxes optimize API, threads, webhooks, forwarding, and MCP—while humans still approve sends.
Discuss this post in AI
Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.
Human mail products compete on unread counts, keyboard shortcuts, and beautiful reading panes. Those are real needs—for people. Agents do not open a reading pane. Agents need an API that creates inboxes, returns threads, fires webhooks, extracts attachments, and sends replies under policy. Pattern Automation built Agent Inbox for that second world.
Optimize for the runtime
An agent inbox prioritizes:
- Provisioning via Python, TypeScript, cURL, and MCP
- Stable addresses per role
- Thread history as structured data
- Forwarding in and out as a core verb
- Ask-friendly draft objects for Neuro OS
Unread badges are irrelevant. Idempotent event delivery is not.
from agentinbox import AgentInbox
client = AgentInbox()
inbox = client.inboxes.create(username="role", domain="agentinbox.space")
Humans still matter—on the gate
Building for agents first does not mean building past people. Outbound mail defaults to Ask. Operators review drafts, approve forwards, and own the role. The human experience is an approval surface and an exception inbox, not a clone of a consumer mail client.
Why the order matters
If you start from a human client and “add API access,” you inherit metaphors that fight the agent: folders as the only taxonomy, manual reading as the happy path, OAuth to a person as identity. Starting from the agent flips the defaults: identity is the role mailbox, reading is programmatic, and people intervene when policy says so.
Neuro OS alignment
Neuro OS is git-backed roles, sandboxes, any approved model, connectors brokered server-side, Ask by default, self-host when required. Agent Inbox is the email channel that matches that architecture. Together they reject the idea that automation must impersonate a person’s Gmail to be useful.
Vision companions: the next trillion users, why AI agents need email, Introducing Agent Inbox.
We do not build inboxes for humans first because agents are the new users of the channel. Humans remain the governors of what leaves.
What human-first would have optimized wrong
Search-as-you-type across personal archive metaphors, snooze, emoji reactions, and inbox-zero gamification do not help a sandbox role. They consume roadmap that should go to webhook reliability, attachment extraction quality, MCP tool clarity, and tenant-safe forwarding rules.
Human-first also pushes OAuth-to-person as the onboarding path because that is how consumer clients grow. Agent-first onboarding is API keys, domain auth, and a first role on Neuro OS with Ask enabled.
Compatibility without capture
People will still read escalations in ordinary clients when an agent forwards to them. That compatibility is intentional. Capture would mean forcing every human interaction through a proprietary reading pane before work can proceed. We refuse that path: the agent inbox is for agents; human tools remain available at the gate.
MCP and SDK as the interface
Agents already speak tools. Exposing inbox operations through MCP and thin SDKs matches how Neuro OS roles work: call a tool, receive structured thread data, propose a send, wait. A pixel-perfect human client would not make that loop clearer. Documentation, examples, and idempotent APIs will.
Agent Inbox gives each role a mailbox people can reply to, with forwarding into Neuro OS when a human must see the thread. Outbound mail defaults to Ask. Run the role on Neuro OS. To scope the first inbox, get started.