AgentMail

AgentMail is an email API for AI agents: create inboxes, send and receive messages, manage threads, and handle webhooks programmatically. Source: AgentMail docs, 2026-07-01

What It Is

AgentMail gives agents real two-way inboxes instead of treating email as an ad hoc SMTP side channel. The docs expose inboxes, messages, threads, drafts, labels, lists, attachments, pods, permissions, webhooks, WebSockets, IMAP/SMTP, custom domains, deliverability, idempotency, and examples for event-driven, sales, labeling, and auto-reply agents. Source: AgentMail docs index, 2026-07-01

The agent-facing surface is broad: official docs are Markdown-readable via .md URLs and /llms.txt; the hosted MCP server is https://mcp.agentmail.to/mcp; the docs also list official skills, CLI, Google ADK, LangChain, OpenClaw, Replit, LiveKit, Sim.ai, x402, and MPP integrations. This makes AgentMail a communication substrate for agent workflows, not just an email-sending API. Source: AgentMail docs index and MCP docs, 2026-07-01

MCP and CLI Surface

The hosted MCP server exposes 17 tools across inboxes, threads, messages, drafts, and attachments. Claude Desktop and Claude.ai use OAuth through console.agentmail.to; Cursor, Windsurf, and similar clients can pass an API key via ?apiKey= or an x-api-key header; Claude Code can add the server with claude mcp add --transport http agentmail https://mcp.agentmail.to/mcp. Source: AgentMail MCP docs, 2026-07-01

This is a good MCP as the Integration Standard example because email has state, auth, side effects, and human-visible consequences. Treat the MCP route as a production integration candidate only after choosing permission scopes, allowlists, webhook verification, dry-run behavior, and human approval for outbound sends. For shell-heavy workflows, prefer the official CLI or SDK when they provide a smaller, auditable command surface.

Email-to-Claude Code Gateway

mvanhorn/agentmail-to-claude-code shows the agent-operator version of AgentMail: email becomes a remote control plane for Claude Code on a trusted Mac. The daemon subscribes to an AgentMail inbox over WebSocket or polls a Primitive inbox, rejects unallowlisted or DKIM/SPF-failed messages, downloads attachments to local paths, writes the email body and attachment paths into a prompt file, then opens a fresh Claude Code session in cmux or Ghostty. If terminal dispatch fails, the message stays unread and the sender gets a bounce, so a task is not silently lost. Source: GitHub README, 2026-07-02

The useful pattern is "inbox as durable task trigger", not email as a chat transcript. The repo's README is explicit about the safety boundary: CC_ALLOWED_FROM controls who can drive the machine, DKIM/SPF failures are dropped, credentials live in env files, and --dangerously-skip-permissions only belongs on a trusted machine and inbox. Attachments are passed as file paths so screenshots/files can reach Claude without embedding raw bytes in the agent's prompt context. Source: GitHub README, 2026-07-02

This belongs next to Claude Code Dynamic Workflows when the trigger needs to originate outside the terminal: phone shortcuts, another agent, cron, email, or a human sending a structured task. Keep it out of default routing until the allowlist, launchd ownership, terminal backend, and audit/logging story are reviewed for the target machine.

Pricing and Startup Program

The reviewed X bookmark announced the Startup Program: three months free on the Startup tier for early-stage AI-agent startups that have raised $500K to $10M and are not current or prior AgentMail customers. The promoted tier includes 150 inboxes, 150,000 emails/month, 150 GB storage, 150 custom domains, SOC 2 report access, and Slack channel support. Source: X/@agentmail, 2026-06-16; AgentMail Startup Program, 2026-07-01

Pricing snapshot on 2026-07-01: Free is $0/month with 3 inboxes, 3,000 emails/month, and 3 GB storage; Developer is $20/month with 10 inboxes, 10,000 emails/month, 10 GB storage, and 10 custom domains; Startup is $200/month with the 150-inbox package; Enterprise is custom and adds unlimited inbox availability, bulk discounts, white-label, usage-based pricing, EU cloud, BYO cloud, dedicated IPs, and OIDC/SAML SSO. Source: AgentMail pricing, 2026-07-01

Version and Source Snapshot

Package snapshot on 2026-07-01:

No current npm packages were found at @agentmail/sdk or @agentmail/client during the registry check. Source: npm registry and GitHub agentmail-to, 2026-07-01

Kevin Fit

Use AgentMail when an agent needs a real external communications channel: outbound outreach, inbound triage, threaded conversations, custom-domain inboxes, webhook-driven processing, or human-in-the-loop draft review. It belongs with Tool-Calling Design and Agentic Infrastructure because the hard part is not "send email"; it is permissioning, reputation, auditability, retries, idempotency, and user-visible side effects.

Do not treat the startup promo as the routing default. The durable thing is the API/MCP/CLI surface for agent-owned inboxes; the promo is only pricing context.

---

Timeline

• 2026-07-02 | Added mvanhorn/agentmail-to-claude-code as the concrete inbox-to-Claude-Code gateway pattern: AgentMail/Primitive inbox, sender allowlist, DKIM/SPF gate, local prompt files with attachment paths, and cmux/Ghostty session dispatch. Source: GitHub mvanhorn/agentmail-to-claude-code, 2026-07-02
• 2026-06-16 | AgentMail announced a Startup Program: three months free on the Startup tier, 150 inboxes, 150k emails/month, and real two-way inboxes for agents. Source: X/@agentmail, 2026-06-16
• 2026-07-01 | Deep-reviewed the startup bookmark and promoted AgentMail into a dedicated tool page with docs, MCP, pricing, package, repo, and visual-artifact snapshots. Source: AgentMail docs/pricing/GitHub/npm, 2026-07-01