Anthropic Built Its Own OpenClaw—and OpenClaw Is in Trouble

Claude Code Channels is real, and it looks a lot like what OpenClaw does.

Anthropic has released Claude Code Channels in research preview — a feature that bridges Telegram and Discord messages directly into a running Claude Code session. The setup is straightforward: create a bot via BotFather, install the plugin, configure the token, restart Claude Code with a --channels flag. Messages sent to the bot arrive in the Claude Code session as events. Claude replies back through the same channel. For an always-on setup, you run Claude in a background terminal or persistent server and the agent waits for a ping to spring into action.

The technical backbone is MCP — Anthropic's own open standard — and the Bun runtime for polling. Documentation is live at code.claude.com/docs/en/channels. This is research preview, requires Claude Code v2.1.80 or later, and needs a claude.ai login; API key authentication is not supported.

The VentureBeat headline called it an "OpenClaw killer." The community reaction on X was similarly blunt: "Claude just killed OpenClaw with this update." The comparison is understandable. OpenClaw's defining feature was persistence — a personal AI agent you could message 24/7 from Telegram, Discord, iMessage, or WhatsApp, that would run tasks and message you back when done. Claude Code Channels offers the same core capability, built on proprietary Anthropic infrastructure.

But the framing inverts the actual dependency graph.

OpenClaw — originally called Clawdbot — was built on Anthropic's Claude model. Steinberger then got a trademark C&D from Anthropic forcing the rename to OpenClaw. He was subsequently hired by OpenAI. So: Anthropic built Claude, Anthropic's model powered OpenClaw's rise, Anthropic's legal pressure pushed the founder to OpenAI, and now Anthropic is shipping the feature that matches OpenClaw's core innovation.

The real story isn't that Anthropic killed OpenClaw. It's that the persistent, asynchronous, messaging-accessible agent paradigm — the thing that made OpenClaw culturally and technically significant — is now being reimplemented by the model provider that seeded it. Whether that's competitive convergence or ecosystem integration depends on your priors.

For OpenClaw users, the practical difference is real: Claude Code Channels is proprietary, tied to Anthropic subscriptions (Pro, Max, Enterprise). OpenClaw is MIT-licensed and runs locally. But Claude Code Channels inherits Anthropic's tier-one model quality and ships with a "Fakechat" demo for local testing. The community sentiment — that Anthropic just made the local-first agent framework obsolete — is a bet on convenience winning over control.

OpenClaw's response will be instructive to watch. Steinberger is at OpenAI now. The protocol layer (MCP) that Anthropic open-sourced is the connective tissue. The community can build channels for Slack, WhatsApp, and whatever comes next. The interesting infrastructure story is who controls the agent runtime, not who ships the UI.