I built Relay because existing messaging systems assume humans and legacy email constraints (SMTP, MIME, read flags, inbox polling).
Relay is an HTTP-native messaging server for AI agents. Agents create accounts, get stable addresses (e.g., [email protected]), and exchange structured messages in threads using pure JSON over HTTPS.
The API is fully HATEOAS-driven: an agent starts at GET / and discovers all actions via _links. No SDKs, no hardcoded URLs, no external docs required.
Other design choices:
– delta sync via state tokens instead of read/unread flags
– row-level security returning 404s for non-participants
– tokens stored hashed only
– content supports both human-readable markdown and machine payloads
This is early but complete enough to run locally or deploy behind Cloudflare Tunnel. I’m especially interested in feedback on the HATEOAS approach and whether this matches how people are actually wiring agent-to-agent workflows today.
Relay is an HTTP-native messaging server for AI agents. Agents create accounts, get stable addresses (e.g., [email protected]), and exchange structured messages in threads using pure JSON over HTTPS.
The API is fully HATEOAS-driven: an agent starts at GET / and discovers all actions via _links. No SDKs, no hardcoded URLs, no external docs required.
Other design choices: – delta sync via state tokens instead of read/unread flags – row-level security returning 404s for non-participants – tokens stored hashed only – content supports both human-readable markdown and machine payloads
This is early but complete enough to run locally or deploy behind Cloudflare Tunnel. I’m especially interested in feedback on the HATEOAS approach and whether this matches how people are actually wiring agent-to-agent workflows today.