Anthropic Ships Its OpenClaw Rival, Connecting Claude Code to Telegram and Discord

Anthropic released Claude Code Channels, a research-preview feature that lets developers send messages to a running Claude Code session from Telegram and Discord. Type a prompt on your phone. Claude runs it on your laptop. The system uses MCP-based plugins to push events into a live local session and send replies back through the same chat app, creating a two-way bridge between messaging platforms and a developer's working environment. Channels requires Claude Code v2.1.80 or later and a claude.ai login, arriving less than two months after OpenClaw creator Peter Steinberger joined rival OpenAI.

The architecture runs on a simple inversion. A channel is an MCP server that runs locally alongside Claude Code as a plugin-backed process. Telegram's plugin, written in Bun, connects to the Bot API and watches for incoming messages. A new message gets wrapped as a <channel> event and dropped into the active session. Claude reads the event, executes whatever work the message requires, and calls a reply tool that sends the response back to Telegram or Discord. The terminal shows the inbound message and a confirmation. The actual reply only appears on the external platform. Events only arrive while the session is open, so always-on setups require Claude Code running in a background process or persistent terminal.

That "push, not pull" distinction matters more than it sounds. Traditional MCP tools sit idle until Claude invokes them. Channels flip the relationship. External systems fire events into the session the moment they arrive, whether that's a chat message from your phone, a CI/CD failure webhook, or a monitoring alert from Datadog. Claude maintains session state across events instead of rebuilding context each time someone opens a terminal window.

Anthropic's own documentation lists CI results, chat messages, and monitoring events as intended use cases. A test suite that fails at 2 AM could fire a webhook into a channel server listening on a local HTTP port. Claude receives the payload, analyzes the error, and in theory could fix the code or send a summary back through the messaging app. Every scenario follows the same sequence. External event arrives. Channel server wraps it as a notification. Claude acts on it, optionally replies. Whether Claude handles the problem well depends on the task, but the plumbing to get the event into the session is what Channels provides.

What a Mac and a phone can do together

MacStories editor John Voorhees spent a few hours testing the Telegram integration on the night it shipped. His results paint a concrete picture of where Channels already works and where it stalls.

Voorhees built and ran an iOS project from his iPhone using xcodebuild, then deployed the app wirelessly to the same device he was chatting from. He compiled a list of 83 saved Readwise articles through the CLI. He kicked off a podcast transcription skill that processed audio, cleaned up the text, and delivered TXT, SRT, and Markdown files back to his phone. All from Telegram, all running on a Mac he wasn't touching.

The friction point showed up fast. Claude Code's permission system has no remote equivalent. If the agent needs approval for a file operation or a shell command, the session pauses until you walk back to your Mac and tap "allow" in Terminal. The workaround is --dangerously-skip-permissions, a flag whose name accurately describes the tradeoff. Voorhees used it. Most serious users probably will too.

Voorhees described setup as "a little fiddly," with one stumble: Bun wasn't installed. Anthropic acquired the JavaScript runtime in December 2025, and Channels plugins require it. Once Bun was in place, the rest followed the documentation. Create a bot through BotFather. Install the plugin. Paste the token. Restart with --channels. A pairing code locks the bot to your Telegram user ID. After that, you're live.

Telegram accepts images. Useful if you're debugging the app on the same phone you're chatting from. Voice messages? Not yet. And if Claude Code needs you to make an actual decision mid-task, the session stalls. Those limitations are real. But the list of things that did work on night one, building iOS apps, running CLI tools, processing audio, suggests the feature arrived closer to useful than most research previews manage.

The OpenClaw problem Anthropic just solved for itself

Steinberger started OpenClaw in November 2025. In interviews, he has said the first version took about an hour: hook WhatsApp to Claude Code, send a message, get a result back. That was the whole thing.

The project grew fast. By February 2026, Steinberger told interviewer Peter Yang that it had reached roughly 300,000 lines of code and supported most major messaging platforms. In that same conversation, he described using OpenClaw to fix a GitHub bug from his phone during a trip in Morocco, sending a photo of a tweet to WhatsApp and watching the agent check out the repository, commit a fix, and reply to the original poster on Twitter.

OpenClaw surpassed 200,000 GitHub stars after going viral in January, even as AI agents broadly continued to stumble on basic tasks that their marketing decks promised they'd mastered. OpenClaw's appeal came with a far looser security posture and broader privileges than Anthropic is now advertising for Channels, a tradeoff that spawned multiple safety-focused forks and alternatives. Anthropic sent Steinberger a cease-and-desist over the original project name "Clawd," then watched from the sidelines as OpenAI hired him in February.

Channels absorbs the core idea and wraps it in Anthropic's security model. Every channel plugin maintains a sender allowlist. Unpaired users get silently dropped. The --channels flag must be explicitly set per session. Configuration lives on the developer's own machine, not on a remote server. On X, AI YouTuber Matthew Berman summed up the reaction in four words. "They've BUILT OpenClaw." Plenty of developers agreed.

That framing oversimplifies. OpenClaw books flights, controls smart home devices, manages social media campaigns across multiple platforms. Steinberger called it "unshackled ChatGPT" in interviews, a phrase that worked as both selling point and warning label. Channels does one thing: two-way chat between a developer and their local Claude Code session. Narrower scope, tighter security posture, clearly different ambition. For now.

Enterprise access and the security tradeoff

Anthropic's enterprise controls read as deliberately cautious, shaped by a company that watched OpenClaw's security reputation closely and decided to overcorrect. Team and Enterprise organizations have Channels disabled by default. An admin must explicitly flip channelsEnabled in managed settings before any user on the org can activate the feature. Pro and Max individual subscribers get access out of the box, but still must opt in per session.

Being registered in .mcp.json alone is not enough to push messages. A channel server must also be named in --channels. That double-gating means an MCP server someone installs for its tools cannot suddenly start injecting events into your session without the developer's knowledge.

A likely enterprise concern is auditability. If Claude Code autonomously commits code, runs tests, and deploys a fix based on a Telegram message, who reviews the work? The permission system offers one answer, but it breaks the mobile use case entirely. The dangerous-skip flag offers another. Neither is satisfying for a compliance team staring at a SOC 2 checklist.

The economic model remains unclear. Claude Code subscriptions were designed for interactive terminal use, not always-on background agents processing messages around the clock. Anthropic hasn't addressed how pricing adapts if Channels graduates from research preview to production feature. That's the kind of quiet question that determines whether enterprises actually adopt this or just admire it from a distance.

What Channels doesn't do yet

The research preview restricts plugins to an Anthropic-maintained allowlist. Right now that means Telegram, Discord, and a localhost demo called Fakechat. No Slack. No WhatsApp. No iMessage. No Microsoft Teams. Developers who want to build custom channels must test with a flag that starts with "dangerously," and those custom channels won't work for anyone else until Anthropic reviews and lists them in the official marketplace.

Early testers have also reported stability issues with the related Remote Control feature, including disconnects and reconnection problems. Documentation has gaps. And Channels requires a claude.ai login, meaning API key authentication and Anthropic Console users are locked out entirely. That's a significant barrier for developers who run Claude Code through organizational API access rather than personal subscriptions.

Voorhees raised a point that extends well beyond developer tooling. Apple has no mobile version of Xcode. Channels effectively turns a Mac into a headless build server you can operate from your phone through Telegram. If coding agents keep filling gaps Apple leaves open, the pressure to ship mobile development tools grows with each release like this.

Boris Cherny, the Anthropic engineer who leads Claude Code development, indicated on X that the company sees Channels as the hackable, developer-focused counterpart to Remote Control. Where Remote Control offers a polished mobile interface for continuing sessions from a phone, Channels exposes the protocol layer that developers can wire to whatever system they want.

And the MCP foundation keeps expanding underneath all of it. Anthropic handed MCP to the Linux Foundation's Agentic AI Foundation last December. The ecosystem crossed 10,000 published servers and keeps growing. Cursor IDE uses it in production. Because Channels is built on MCP, the community can build connectors for any platform with an API without waiting for Anthropic to ship them. The plugins for Telegram and Discord already sit on GitHub under Anthropic's official repositories, open for inspection and contribution.

Channels won't replace OpenClaw for users who want an unrestricted AI agent managing their entire digital life from a phone. But for developers who want to check on a build, kick off a test run, or ask their codebase a question from the grocery store, the terminal just stopped being the bottleneck. Anthropic's bet is that controlled access to a running session beats unconstrained access to everything. Whether the market agrees depends on how fast the plugin ecosystem grows, and how many developers decide that --dangerously-skip-permissions is a flag they're comfortable typing every morning.

Perplexity Launches Computer, an Agent Platform Orchestrating 19 AI Models at Once

Perplexity on Wednesday unveiled Computer, a multiagent orchestration system that routes tasks across 19 frontier AI models to handle end-to-end workflows from research and design through code deploym

The Implicator

OpenAI Launches Frontier to Manage AI Agents From Rival Vendors in One System

OpenAI on Thursday released Frontier, a platform that lets companies build, deploy, and manage AI agents from multiple vendors inside a single system. The product works with agents built by OpenAI, by

The Implicator

OpenAI Launches Codex Desktop App for macOS With Multi-Agent Workflows and Doubled Rate Limits

OpenAI released a macOS desktop app for Codex today, turning its AI coding agent into a standalone application that can run multiple agents across different projects at the same time. The company also

The Implicator

Marcus Schuler

San Francisco

Tech translator with German roots who fled to Silicon Valley chaos. Decodes startup noise from San Francisco. Launched implicator.ai to slice through AI's daily madness—crisp, clear, with Teutonic precision and sarcasm. E-Mail: [email protected]