Anthropic Ships Claude Code Channels, an OpenClaw Competitor

Anthropic has launched Claude Code Channels, a research-preview feature that lets outside systems push messages into a running Claude Code session. VentureBeat framed it as an OpenClaw killer. After a proper pass through the documentation, my read is that the headline oversells the overlap, and the feature is more interesting for what it deliberately leaves out.

What Channels Actually Does

A channel is an MCP server that pushes events into the Claude Code session you already have open, so Claude can react to things that happen while you are away from the terminal. Anthropic’s docs describe two shapes. A chat bridge is two-way: you text Claude from your phone, the work runs on your machine against your real files, and the answer comes back in the same chat. A webhook receiver is one-way: your CI pipeline or error tracker POSTs to a local port, and the alert lands where Claude already has your files open and remembers what you were debugging.

That “already open” part is the design decision that matters. Events only arrive while a session is running, so an always-on setup means keeping Claude alive in a persistent terminal or a background process. Channels do not spawn fresh cloud sandboxes the way Claude Code on the web or Claude in Slack do. They feed the local session you were sitting in before you walked away.

What Is Actually in the Preview

The research preview ships three real platforms, Telegram, Discord, and iMessage, plus fakechat, a localhost demo you can try without creating a bot anywhere. Each one is a plugin, and the pre-built plugins are Bun scripts, so you need Bun installed before any of this works.

Setup is a plugin install rather than a new subcommand. For Telegram it looks like this:

/plugin install telegram@claude-plugins-official
/reload-plugins
/telegram:configure <token>
claude --channels plugin:telegram@claude-plugins-official

You need Claude Code v2.1.80 or later, authenticated through claude.ai or a Console API key, and channels are not available on Amazon Bedrock, Google Cloud’s Agent Platform, or Microsoft Foundry. Pro and Max users can opt in per session with the --channels flag. On claude.ai Team and Enterprise plans, nothing arrives until an org Owner flips the channelsEnabled setting, and admins can pin an allowlist of approved plugins on top of that.

The Security Model Is Most of the Feature

An open inbound channel into a coding agent is a prompt-injection vector, and the docs treat it that way. Every approved plugin keeps a sender allowlist, and messages from anyone not on it are silently dropped. Telegram and Discord bootstrap the list with a pairing code: you DM your bot, it replies with a code, and you approve that code inside your Claude Code session. iMessage trusts texts from your own Apple ID automatically and makes you add anyone else by handle. Being registered in .mcp.json is not enough on its own, either; a server also has to be named in --channels for that specific session.

The newest piece is permission relay, which needs v2.1.81 or later. A channel can forward Claude Code’s tool-approval prompts to your chat app, so you can reply “yes” or “no” with a short request code from your phone and the local dialog closes. The docs are blunt about the trade: anyone who can reply through the channel can approve tool use in your session. Keep that allowlist to people you would hand your laptop.

How It Compares to OpenClaw

OpenClaw, sitting above 380,000 GitHub stars, is a model-agnostic gateway that connects to a long list of messaging platforms and runs as its own always-on assistant. Channels is three platforms and a demo, gated behind an Anthropic-curated allowlist, in a research preview whose flag syntax may still change. During the preview, even a channel you build yourself needs a --dangerously-load-development-channels flag to run locally.

My verdict after reading both docs sets: these tools want different jobs. OpenClaw wants to be the assistant that lives in your chats. Channels wants your existing working session, with its hooks and MCP servers and accumulated context, reachable while you make coffee. The bet Anthropic made is a small approved surface with an unusually careful permission story, which is the choice you would expect after a year of MCP supply-chain incidents.

Getting Started

Start with fakechat before you wire up a real platform. Install it with /plugin install fakechat@claude-plugins-official, restart with the --channels flag pointing at it, and a chat UI appears at localhost:8787 that pushes straight into your session. If it feels useful, graduate to Telegram or Discord with a bot token, or iMessage if you live on macOS and can grant Full Disk Access. Our Claude Code guide covers the base setup if you are not running it yet.

Correction (July 2026): an earlier version of this story described Channels as a Max-only feature spanning 20+ platforms, enabled with claude channels add. The research preview works as described above: three platforms plus a demo, installed as plugins and enabled per session with --channels.