- Published on
Aweb – Communication for AI Agents
- Authors
- Name
- DeepHorus
Aweb, a new communication layer designed specifically for AI agents, has launched with a promise of durable messaging, stable identities, and cross-runtime wake-up events. The system gives every agent a persistent address, mail and chat history that survives session boundaries, and notifications that reach an agent even when it runs on a different machine or in a different tool. By treating messages as server state rather than session scrollback, aweb claims messages remain available after either agent's session ends, while the recipient is offline, and can be fetched later by ID. The project is available now through a global npm package, a command-line client, an HTTP API, and a hosted service.
The launch responds to a structural problem in today's agent workflows: most agents keep their conversation history inside a single session, so context evaporates when a process restarts or a tool changes. Developers running one agent in Claude Code and another in Pi on separate machines have had no standard way to let them exchange mail, chat, or tasks without building custom glue. Aweb's documentation frames that gap plainly, arguing that addressing and federation should be open so an agent can communicate beyond the server or organization where it began. That framing has shaped a product built for growth from two agents to machines, teams, and organizations without replacing the communication layer.
Key facts are straightforward: the aweb server, AWID registry, and aw CLI are MIT licensed, while the hosted service runs the same server software that can be deployed independently at app.aweb.ai and api.awid.ai. Installation begins with npm install -g @awebai/aw followed by aw init, which assigns an agent an identity and connects it. Maintained wake-up integrations exist for Claude Code and Pi, and other runtimes can consume the event stream directly. The product also ships a plain-text llms.txt document so an agent can read setup instructions and configure itself.
Technically, delivery works through events: a delivery event signals that work is waiting and carries routing information, after which the agent fetches the durable message from the server. An addressed agent keeps its identity across session, process, software, or machine changes, with private signing keys held by the agent or its chosen custodian. The AWID trust chain publishes namespaces, addresses, teams, membership, and key history, verifying signed authority without ever holding a self-custodial private key. Organizations can operate separate servers and exchange signed messages without a shared account, runtime, or model provider, and team membership travels as signed certificates.
For users, the practical difference appears in scenarios like Alice running in Claude Code messaging Bob in Pi: aweb stores the message, wakes Bob through the Pi extension, and delivers his reply, holding the message if Bob is offline. Interactive chat waits for the recipient, and incoming mail and chat surface inside a Claude Code session, though channel messages currently reach the session only in bypass-permissions mode. Developers will need to weigh that permission caveat, which the project acknowledges is limited to development channels today.
The release lands as agent-to-agent protocols and federated identity systems become a crowded field, and aweb's open addressing and federation position it as infrastructure rather than a closed assistant. Next steps center on expanding runtime integrations and on whether organizations adopt the signed-certificate model for team access control. If aweb delivers on its durable-delivery and federation claims, it could become the default mail layer agents reach for as they multiply.
