The agent ecosystem is finally getting a dedicated communication layer that doesn't pretend to be human email. Crabmail, launched by 23blocks, introduces the open Agent Messaging Protocol, allowing AI agents to exchange signed messages via addresses like scout@yourteam.crabmail.ai. This isn't SMTP; it's a purpose-built relay for machine-to-machine interaction, currently free for a single agent.
Cryptographic Trust Over Legacy Standards
The core innovation here is cryptographic verification. Agents generate their own Ed25519 keypairs, retaining the private key while Crabmail stores only the public counterpart. This ensures that messages can be verified as authentic, with untrusted or badly signed payloads automatically wrapped as data rather than executed instructions. The protocol is transparent, with provider public keys and version numbers accessible via a simple curl request to api.crabmail.ai/.well-known/agent-messaging.json.
Integration Without the Password Hell
Onboarding is stripped of friction. Users sign up via email codeβno passwords requiredβand hand a user key to their agent. The agent then self-registers, creating its own identity and beginning communication immediately. It supports major players like Claude Code, Codex, and Grok, provided they can call a REST API. While Pro ($9/month) and Team ($29/month) tiers exist, they are not yet open for purchase, keeping the current focus on the free single-agent tier.
Key Takeaways
- Crabmail is strictly for agent-to-agent messaging and cannot receive standard internet email.
- Messages are signed with Ed25519, and agents hold their own private keys for security.
- The plugin is open-source, and the protocol is defined at agentmessaging.org.
- Cross-provider reach is not yet verified, limiting current utility to the Crabmail network.
The Bottom Line
This is the infrastructure patch we needed for the swarm era. Stop trying to force LLMs through SMTP gates and give them a protocol that speaks their language: signed, structured, and stateless.