Claude Code Channels Lets You Control Your Dev Environment From Telegram and Discord

Anthropic just shipped one of the most requested features in the AI coding agent space: the ability to talk to Claude Code when you’re not sitting at your terminal.

Claude Code Channels, released in v2.1.80 on March 19, 2026, turns messaging platforms into remote interfaces for active Claude Code sessions. You DM a Telegram bot or a Discord bot from your phone, and your message lands directly in a running Claude Code session on your machine. Claude reads it, acts on it, and replies back through the same platform.

Open post on X

Thariq Shihipar from Anthropic’s Claude Code team announced the feature, which launches with support for Telegram and Discord, plus a localhost demo called “fakechat” for testing.

What Channels Actually Does

The core idea is simple: push external events into a running Claude Code session so Claude can react to things while you’re away from the keyboard.

That means two main use cases. First, you can use your phone as a remote control. You’re on the couch, in a meeting, on the train. You message your Telegram bot: “run the test suite and tell me what fails.” Claude gets the message, runs the tests, and sends back the results to your Telegram chat.

Second, you can pipe external events into Claude’s context. CI failures, monitoring alerts, webhooks from third-party services. Anything that can send an HTTP request can, through a channel server, push an event into Claude’s session and let it respond.

Channels can be one-way (events flow in, Claude reacts) or two-way (Claude reads the event and replies back through the same platform). The Telegram and Discord integrations are both two-way.

How It Works Under the Hood

A channel is an MCP (Model Context Protocol) server that runs as a subprocess alongside Claude Code. The architecture is clean: Claude Code spawns the channel server, communicates with it over stdio, and the server bridges between the external platform and the session.

The channel server declares a claude/channel capability and emits events as MCP notifications. These arrive in Claude’s context as XML tags:

<channel source="telegram" sender="michael">deploy the staging branch</channel>

For two-way communication, the server exposes a standard MCP tool (like reply) that Claude calls to send messages back to the originating platform.

To start a session with channels enabled:

claude --channels plugin:telegram@claude-plugins-official

The setup for Telegram takes about two minutes. Create a bot via BotFather, install the plugin, configure the token, and pair your sender ID. Discord follows a similar pattern through the Discord Developer Portal.

Security Model

Anthropic clearly thought about the attack surface here. Every channel maintains a sender allowlist. Only paired and approved sender IDs can push messages into your session. Unpaired senders are silently dropped.

This matters because an ungated channel would be a prompt injection vector. If anyone could send messages to your Claude Code session through a public Telegram bot or Discord server, they could instruct Claude to do things to your codebase. The docs call this out explicitly and recommend restricting access to known sender IDs only.

For Team and Enterprise orgs, channels are disabled by default. An admin has to explicitly flip the channelsEnabled setting before anyone on the team can use them.

One limitation worth noting: channels only work with claude.ai login authentication. Console and API key auth are not supported. This means channels run against your subscription quota (Pro, Max, Team, Enterprise), not API billing.

The OpenClaw Comparison

It’s impossible to talk about Channels without talking about OpenClaw. The open-source project pioneered exactly this pattern: connecting AI coding agents to messaging platforms so developers could interact with them remotely. OpenClaw supports Telegram, Discord, WhatsApp, Slack, and more, and it works with Claude Code, Codex CLI, and Gemini CLI.

VentureBeat called Channels an “OpenClaw killer,” which is a strong framing but not wrong directionally. Anthropic has taken the most popular pattern from the community and built a first-party version with tighter security guarantees, simpler setup, and no need for dedicated hardware.

OpenClaw still has advantages. It supports more platforms (WhatsApp, Slack, Gmail), works with multiple AI coding agents (not just Claude Code), and has a more mature plugin ecosystem. But for developers who primarily use Claude Code and want Telegram or Discord access, Channels removes the need for a third-party tool entirely.

Ethan Mollick, the Wharton professor who has become one of the most cited voices on AI tooling, noted the strategic implications:

Open post on X

His point is worth sitting with. Anthropic is shipping features at a pace where community innovations get absorbed into the official product within weeks. That’s a strong position for Claude Code users, but a challenging one for independent tool builders.

What You Need

Claude Code v2.1.80 or later
A claude.ai subscription (Pro, Max, Team, or Enterprise)
Bun runtime for pre-built plugins
A Telegram bot token (via BotFather) or Discord bot (via Developer Portal)

The feature is in research preview, so the --channels flag syntax and protocol may change. Anthropic is collecting feedback before locking down the API.

Building Custom Channels

The channels reference docs lay out everything you need to build your own channel server. The protocol is straightforward MCP with a custom capability declaration. You can use Bun, Node, or Deno.

Custom channels need the --dangerously-load-development-channels flag during the research preview since they’re not on the approved allowlist. Once you’ve built something stable, you can submit it to the official plugin marketplace for security review and inclusion.

The official plugin source code for Telegram and Discord is public, which makes it a solid starting point for anyone building a Slack, iMessage, or webhook-based channel.

Why This Matters

The shift from “sit at your terminal and type” to “message your agent from anywhere” has been coming for months. OpenClaw proved the demand. Anthropic shipping it as a first-party feature confirms that remote, asynchronous interaction with coding agents isn’t a novelty. It’s becoming a standard part of the workflow.

The practical impact is straightforward. You can start a Claude Code session on your workstation, leave for lunch, and keep directing it from your phone. You can set up a CI pipeline that pushes test failures directly into Claude’s context and lets it propose fixes. You can run a persistent session on a VPS and interact with it from anywhere.

None of this was impossible before. OpenClaw and various custom setups have done it for months. What Channels does is make it easy, official, and secure enough that you don’t need to think about the plumbing.

Try it at code.claude.com/docs/en/channels.

Sources: