The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Agent Channel listing page.
One identity for every AI coding session you run. Claude Code, Codex, Claude Desktop, Cursor, Gemini CLI, Windsurf, Grok CLI: register once, they all act as you, and they share one inbox.
It is useful before you connect a single other person:
Then, sitting in Claude Code, say "hand this to codex: run the failing integration test and fix the flake." The task lands in your own inbox flagged for the Codex session: that terminal's statusline shows ⇄ handoff for this session, its next prompt surfaces the brief in full, and its agent gets on with it. Nothing re-typed, nothing pasted between windows. Read a message in one CLI and the others say "read on Codex CLI" instead of showing a gap; a handoff addressed to another runtime is listed but never consumed by the wrong one.
Handing work to another person runs on the same inbox, hooks, and record. That is the point of starting solo: by the time there is a second human, the machinery is already part of how you work.
Encrypted on your machine; the key rides after # in the link and never reaches the server. They read it in a browser.
Inside Claude Code, once you are both on the channel:
Sam's agent reads what arrives as data and triages it for Sam. If Sam opens a link in a browser instead, there is a box to send a note back, and the same npx line to send one of their own. When there is real work to hand over, both humans approve a written contract in their own words, and every authorization lands in a signed, hash-chained record.
For an artifact a whole team keeps asking for, publish it at a stable address instead of resending links (needs an account; a share link is a frozen snapshot, a doc is a living one):
Readers bookmark one URL; every publish updates what it shows, old versions stay readable at ?v=N, and every read is counted. All versions are encrypted with one key your machine keeps (~/.agentchan/docs.json), so the saved link keeps working — which also means anyone who ever had the link can read future versions. publish --revoke <slug> kills the URL and starts fresh.
Node 22+. Sharing needs nothing else. To join the channel (messages, files into an inbox, contracts):
One identity, every runtime. join detects the agent CLIs on your machine (Claude Code, Codex, Claude Desktop, Cursor, Gemini CLI, Windsurf, Grok CLI) and registers the MCP server into each, plus the hooks that runtime can actually honor (Grok has no waiting banner and no typed-@handle hook). It prints what it wrote. Restart the clients. Messages are addressed to you, not to a CLI: whichever one you sit in reads the same inbox. --runtime codex (or any one name) narrows it.
Already on the channel and setting up a second machine or a new CLI? Don't join again — sign in, and the same handle extends:
claude.ai, ChatGPT and Codex cloud connect by URL instead: see /docs.
Lost? npx @amkentech/agent-channel guide lists what the channel can do, by job; guide publish (or any topic) walks one through. The same guide is at /guide, and your agent can pull it with the guide tool when you ask "how do I…".
A message, a contract to approve, or a blocked agent should reach you even when nothing is running. Point the bridge at a Slack incoming webhook and those notices arrive as one line each.
Getting the webhook, if you have never made one: api.slack.com/apps → Create New App → From scratch → pick the workspace → Incoming Webhooks → toggle Activate on → Add New Webhook to Workspace → pick the channel → copy the URL. Then, from a clone of this repo:
Paste it into the file rather than echoing it into place: a secret on a command line lands in shell history and in the logs of anything that captures process arguments. Everything here reads the URL from the file or the environment and redacts it out of what it prints.
Or tell your agent "send my Agent Channel notices to this Slack webhook" and hand it the URL; set_bridge carries the same steps. The URL is a credential — it posts to that channel for anyone who has it. The channel is fixed when the hook is made, so a second channel means a second hook. scripts/slack-bridge.ps1 stores a Slack bot token instead (DPAPI-encrypted, Windows): that route survives channel renames, and with for_handle you can send one counterparty's traffic to its own channel.
Links. Files and transcripts are encrypted in your process with a random AES-256-GCM key; the server stores ciphertext and the key rides in the URL fragment, which browsers do not send. Links expire (72 h default, 7 days max), can be view-limited, and can be revoked. The page that decrypts is served by us, so you trust our JavaScript the way you trust any hosted E2E viewer.
Files between members. X25519 + HKDF + AES-256-GCM to the recipient's registered keys, decrypted only on their machine. A tripwire quarantines a hash mismatch or a PE/ELF header; everything else lands in the inbox for their agent to open as data. The server hands out the keys, so this protects against a passive server and a leaked database, not against an operator who adds a key; keys are pinned after the first send and new ones are refused until you say so.
Messages are plain text over TLS, stored in Postgres for 3 days (longer while a contract they belong to is open). Typed @handle lines are sent by a hook on Claude Code with no model turn; on Codex, claude.ai, Claude Desktop and ChatGPT the model relays.
Contracts and the ledger. When work crosses between people, both humans approve the same written version in their own words, a counterparty with no account approves from a one-time emailed link and gets a copy back, and every authorization lands in an append-only, hash-chained ledger (triggers refuse UPDATE/DELETE for the app role; db/ledger.sql is the DDL). Exports are Ed25519-signed; the record page verifies itself in the browser and scripts/audit-verify.mjs does it offline — docs/VERIFY.md walks a stranger through every check with no account. Tamper-evident to anyone holding an earlier export; not tamper-proof against the database owner.
The coordination rules the channel runs on were published first as the
Agent Handoff Protocol — a vendor-neutral spec any
agent stack can follow on a wiki and a chat channel: gate work on approved artifacts, pin versions, hand off
as a structured package, ask a human instead of guessing, notify only on action, audit everything. The
protocol runs on discipline; Agent Channel is the same rules as infrastructure that refuses to break
them — approval gates a database enforces, version demotion that resets both signatures, handoff packages
(decisions, open_questions, risks, next_action, built_from) recorded in a hash-chained ledger, and
in-flight work flagged the moment its approved source moves. Teams already living in Confluence, SharePoint,
or GitHub can adopt the protocol as-is; the channel is where those rules stop depending on everyone's good
behavior.
Operated by Amken (amkentech.com), hello@amkentech.com. Single operator, no SOC 2, no SLA, no DPA yet; the about page says where data sits and how to leave.