Aweb launches federated, MIT-licensed messaging and wake-up layer for AI agents

Mail that outlives the session

Aweb, an open-source communication layer built for AI agents, reached Hacker News's front page on 1 October, according to the Hacker News listing. The pitch is that software agents need the same basics human collaboration relies on: stable addresses, mail that survives a crashed session, and a way to be woken when work arrives. The aweb server, its AWID identity registry and the aw command-line client are all MIT-licensed, and independently operated servers can federate with one another.

On aweb, a message is stored on the server rather than in a running session's buffer: it remains available after either agent's session ends, while the recipient is offline, and can be fetched later by ID. The worked example on the project's site has an agent named Alice, running in Claude Code on one machine, send mail to an agent named Bob, running in the Pi runtime on another. aweb stores the message, wakes Bob through the Pi extension and routes his reply back to Alice; if Bob is offline, the message waits for his return. A send-and-wait chat mode blocks until the recipient answers, for questions that need an immediate reply. The server also tracks presence and maintains a shared task queue, and the site points to a 95-second unedited recording, dated 5 September, in which a cross-runtime question is answered in nine seconds.

Waking agents across runtimes

The CLI and HTTP API are runtime-independent: any agent that can call the CLI, the HTTP API, MCP tools or the event stream can participate. Maintained wake-up integrations exist for two runtimes today, a Claude Code channel plugin and a Pi extension (installed with pi install npm:@awebai/pi@latest) that wakes Pi when mail, chat or tasks arrive.

The Claude Code path carries a caveat worth noting: channel messages reach the session only when it is started in bypass-permissions mode with development channels loaded, and in the default auto mode messages arrive but are not shown. aweb also publishes an llms.txt file, the same product and setup documentation in plain text, so a user can simply tell an agent to read it and configure aweb itself.

Identity rooted in DNS, not the delivery server

Every agent gets an AWID address of the form domain/name, with a trust chain that starts in DNS: the namespace controller authorizes a team controller, the team controller signs membership certificates, and each agent signs with its own key. Recipients can verify that chain without trusting the communication server to define identity.

Custody and hosting are deliberately separate choices. Agents can hold their own signing keys or use a custodian, and a 'Bring Your Own Team' mode lets an organization run its namespace under its own domain. In that mode, according to the site, the hosted aweb service can deliver a team's messages but cannot add members or manufacture team authority. Recipients also control delivery: accept first contact from any verified sender, or restrict it to verified teammates and explicit contacts.

Self-host or use the hosted service

Both deployment modes use the same CLI and protocol. Self-hosting starts from the project's GitHub repository with Docker Compose: clone the repo, put a generated AWID service token into .env, run docker compose up, and health-check the server and registry on local ports 8000 and 8010. An agent then gets an identity with aw init --new-team. The hosted option runs the identical communication server at app.aweb.ai, with the AWID registry at api.awid.ai. Either way, installation begins with npm install -g @awebai/aw.

Why it matters

Agent-to-agent coordination today is mostly improvised: a human relays context between windows, or state lives in a session log that dies with the process. Aweb's bet is that agents become first-class network citizens, durably addressable, verifiably identified and reachable even while offline, which resembles email-style federation far more than a chat window's history.

Because the stack is MIT-licensed and self-hostable, and because identity authority is anchored in DNS rather than in the operator of any one server, an organization can adopt it without handing trust to a single vendor. The open questions are the usual ones for a young protocol: whether runtimes beyond Claude Code and Pi adopt the event stream, whether the permission-bypass requirement for the Claude Code channel matures into something safer than a development flag, and whether agent-to-agent mail becomes routine plumbing. What aweb offers now is a coherent, inspectable version of the missing pieces, durable queues, wake-up events, verifiable identity and federation, in one package.