Show HN: Agentlink – Messaging for Claude Code agents on different accounts

Let AI agents on different Claude accounts — and different machines — message each other. Claude Code's own messaging needs every session on one account; agentlink doesn't. Address any agent as name@host (e.g. reviewer@laptop.local). An incoming message wakes an idle Claude Code or Codex session, just like agents talking on a single machine. Watch the whole one-minute demo, from setup to reply, at poly.io/agentlink What you need Claude Code 2.1.318 or newer on each machine. Codex 0.167 or newer is woken too, and other MCP agents can join (for a machine without Claude Code, set AGENTLINK_SKIP_CHECKS=1 for the installer). An Apple silicon Mac or Linux. Intel Macs aren't currently supported. Machines on the same network, or Tailscale on each of them. The installer brings everything else. It needs no sudo. First machine — paste into Claude Code (or run the command yourself): Set up agentlink on this machine: run curl -fsSL https://raw.githubusercontent.com/tombaldwin/agentlink/main/install.sh | sh, then show me the prompt it prints for adding other machines, and tell me to type /reload-plugins. Or install the Claude Code plugin, which teaches Claude to set agentlink up and use it, then ask Claude to set up agentlink: /plugin install agentlink --marketplace tombaldwin/agentlink Every other machine — paste the prompt the first machine printed. It looks like: Join my agentlink network: run curl … | sh — then tell me to type /reload-plugins. Each invite adds one machine and expires after 24 hours. The pinned key means curl only talks to the machine that made the invite. That key exists only while the invite is open, so a used invite line stops working (curl reports a key mismatch or a 404). Run agentlink invite on any member for a new one. Then type /reload-plugins in Claude Code. That session is now on the network, and so is every new one. (If that doesn't work, restart it: claude --continue picks up where you left off.) That's it: no config files, no accounts, no server to run. Setup adds agentlink's tools and a small background hook to Claude Code's settings (and Codex's, if it's installed), and agentlink uninstall takes them out again. On Apple silicon Macs it installs agentlink as a signed app (~/Applications/agentlink.app, from the inviting machine or the latest release, refused unless Apple's certificate for agentlink's team signed it); macOS asks once to let agentlink find devices on your local network: click Allow. Elsewhere it uses npm, and brings its own Node.js if the machine doesn't have a recent one. For the next machine, ask Claude on any member to add one (or run agentlink invite). Your first message Agents are named after their project folder, so a session in ~/code/reviewer on the machine desk is reviewer@desk.local. Ask Claude who's online, then ask it to message someone: You, to Claude on your laptop: Ask reviewer@desk.local why the build is failing. Claude: Sent. pushed to reviewer@desk.local's live session The session on desk wakes, even if it was idle, and works on it. Its answer wakes the session on your laptop in the same way. You don't need the exact address. reviewer@desk works (.local is added, and case doesn't matter), and so does a machine's Tailscale name or IP address. You can also just say "ask the reviewer on my desktop": Claude checks who's online and works out which agent you mean. More ways to use it, such as a reviewer on another account, waking a session from CI, or a nightly job that reports in the morning, are in Recipes. What you get Wake on message. A background hook wakes the session when a message arrives, even when it's idle, and Claude reads it and acts on it (it still stops at any permission prompt). Tools: send, peers, inbox, plus network and invite for managing the network. Removing machines and uninstalling stay in the terminal, so you always see the command first. Nothing gets lost. Messages are stored on disk at every hop and retried. If the recipient isn't running, its machine keeps the message; if the whole machine is off, copies wait on up to 3 other online machines and whoever sees it return first delivers. Copies that reach a machine twice are dropped, though a session that dies just as a message is shown gets it again when it restarts. Messages expire after 7 days. A session can have 2,000 unread messages waiting, and a machine's store can reach 1 GB (maxWaiting and maxStoreMB in ~/.agentlink/config.json). Past that, a message from another machine waits at the sender and is retried, and a send from this machine fails with an error, so a runaway agent stops there. Zero-config discovery. Machines on the same LAN find each other automatically (encrypted UDP beacons) and share what they know, so every machine learns the whole network. Multiple Claude accounts. Every CLAUDE_CONFIG_DIR account setup can find is enabled (~/.claude* dirs, shell aliases, and the one the shell is using); run agentlink setup again to add new accounts. Beyond the LAN. With Tailscale on both machines, agentlink invite includes the Tailscale name and joining machines use it automatically. A machine that moves between networks is found again at its other addresses, within a couple of minutes. Codex. If Codex is installed, setup adds agentlink to ~/.codex/config.toml too. The next time Codex starts, it asks you to trust agentlink's hook: trust it, and a message gives an idle Codex session a turn (through codex queue, Codex 0.167 or newer). Codex sessions are named after their project folder. Other agents. Any MCP agent can join with agentlink mcp --poll (it reads with inbox instead of being woken). Channels and managed accounts. If your organisation's policy turns channels off, agentlink claude sessions can't tell, and pushed messages don't show up. Set AGENTLINK_POLL=1 so they're read with inbox instead. Name a session with AGENTLINK_NAME= claude (default: the project folder name). Compared with Claude Code's own messaging Claude Code can already message between your own sessions: on one machine over a local socket, and between machines through Anthropic's servers, with a claude.ai sign-in and Remote Control. If that covers you, you don't need agentlink. agentlink is for sessions on different Claude accounts, agents other than Claude Code, and messages that stay between your own machines. Commands Claude handles everyday use and invites; these are for terminals and scripts. agentlink send agentlink peers agentlink status agentlink invite [--list|--revoke] agentlink remove agentlink compact [--yes] agentlink uninstall agentlink help (or agentlink --help) shows a command's options. agentlink peers says why a machine is offline (the node log, ~/.agentlink/node.log, records when a machine stops or starts answering). Tab completion, including the addresses of agents that are online after agentlink send: add eval "$(agentlink completion zsh)" to ~/.zshrc, or eval "$(agentlink completion bash)" to ~/.bashrc. Upgrading Re-run the install line from Setup on each machine (or npm install -g @tombaldwin/agentlink followed by agentlink setup). It keeps the machine's key, network and messages. Upgrade all your machines: a machine running an incompatible version is reported as such (… runs an older version of agentlink) and can't exchange messages until it's upgraded. Claude sessions that were running during the upgrade say they need restarting (claude --continue resumes the conversation); they never stop the newer node. Security Each machine has its own key. The network is the list of machines' keys, kept and shared by the members. Any member can invite machines, and a machine can remove the ones that joined after it (the first machine can remove anyone). Encrypted, and nothing for strangers. Machines talk TLS and prove their keys to each other inside the connection. A laptop at a café that tries a home address now held by a stranger doesn't say who it is or send anything before the check fails. And anyone who connects to a machine sees only a throwaway key, so a laptop can't be recognised as it moves between networks. Joining. An invite is a one-time token plus a key the inviting machine presents while the invite is open. Someone watching the network learns nothing from a join, and can't swap the installer. A used or expired invite is worthless. Machines can't impersonate each other. Every message is signed by the machine that sent it, and machines relaying a copy can't change it (they can read it: members are trusted). agentlink remove takes a machine off the network; members that were offline learn about it when they come back. Members are trusted. Any member can add machines, and every agent on a member machine can message yours. Only add machines you control. Prompt injection. Messages come from other AI agents. agentlink tells Claude to treat them as requests from a colleague, not instructions from you, and to check with you before anything destructive or sensitive. See SECURITY.md for details and known limitations. Only interactive sessions are woken. claude -p runs and SDK sessions aren't, and nothing about them changes: their messages wait for the inbox tool. A Codex session is woken from its first prompt on, when Codex runs agentlink's hook. After you close a Codex window, Codex's background server can keep the session open, and then answers its messages by itself. Two Codex sessions in the same folder can't be told apart: only the newer one is woken. A Codex turn you stop with Esc holds messages back until your next prompt. A woken session reads messages by itself, but stops at permission prompts as usual, so replying may wait for you. Setup allows only the inbox and peers tools to run unattended. Senders see "pushed to … live session" when the session has been told, not when Claude has read the message. A session idle for more than 20 days stops being woken until its next turn. So does one you interrupt with Esc after it's woken and before it reads its messages; they're still there to read. A VPN that blocks local traffic makes your machines show as offline. Allow local network access in its settings, or use Tailscale. Messages wait until the machines can reach each other. Over SSH on a Mac, Claude Code can't reach its login while the keychain is locked: run security unlock-keychain first. Re-running agentlink setup keeps the machine on its network; pass --new-network to start a new one. A machine's name is fixed when it joins. To rename it, run agentlink remove on a machine that joined before it, then run a new invite line with | sh -s -- --host in place of | sh. Messages already queued for its old name wait until they expire. A removed machine comes back with a new invite: it gets a new key, and its messages, delivered or not, are dropped. Joining another network drops them too. A machine joins once, for one computer login. A second login on the same machine is refused, because the machine's name is already a member. Several Claude accounts under one login are fine. Apple silicon Macs and Linux. Intel Macs aren't currently supported: the npm install runs there, but macOS may stop it reaching other machines (see below), and there's no app for them yet. On Linux, .local names need mDNS (avahi-daemon and libnss-mdns); without it, use the IP line the invite offers. Port 7777 must be open in any firewall between your machines. On macOS, agentlink needs Local Network access to reach your other machines. The app asks for it by name; if you clicked Don't Allow, turn agentlink on under System Settings → Privacy & Security → Local Network. An npm install on a Mac (AGENTLINK_APP=0) runs as plain node at login, which macOS may block without asking: agentlink peers then shows machines as offline with EHOSTUNREACH. Development npm test # unit tests + multi-node end-to-end: live delivery, offline session, offline machine # via holders, dedupe, poll mode, outsiders, forged and altered messages, strangers # at stored addresses, relays, invites and joining, removal, big backlogs, node # restart, IP/Tailscale aliases The macOS app (agentlink as a single signed executable, so macOS asks for Local Network access by name; see scripts/build-app.mjs): npm run build:app # dist/agentlink--macos-.zip npm run build:app -- --arch x64 --sign "Developer ID Application: …" --notarize AGENTLINK_TEST_EXE=dist/arm64/agentlink.app/Contents/MacOS/agentlink npm run test:e2e Releases (signed and notarized app, GitHub Release, npm): see docs/releasing.md.