On March 20, Anthropic launched Claude Code Channels as a research preview — and it directly tackles the biggest gap in the Claude Code workflow: what happens when you step away from your terminal?
The gap that OpenClaw filled
OpenClaw, an open source project launched in November 2025, went viral by solving a simple problem: letting developers message their AI agent from their phone. With 200K stars on GitHub, it became clear the demand was real. Devs wanted to start a task, step away, and get a notification when it finished — all from Telegram or WhatsApp.
Anthropic’s answer is Claude Code Channels: native, first-party support for exactly that workflow. Built on MCP, with more robust security and setup that takes less than five minutes.
What Channels does, specifically
Channels inverts the traditional interaction model. Instead of you going to Claude, Claude comes to you.
A traditional Claude Code session is synchronous: you open a terminal, write a prompt, wait for the response. You close the terminal, the interaction ends. With Channels, Claude Code runs in the background on your machine. External events — a Telegram message from your phone, a CI failure, a monitoring alert — are injected directly into the active session. Claude processes each one with the full context of the project and responds through the same channel.
Anthropic calls it “push, not pull”. The distinction matters more than it seems.
The architecture in a paragraph
Channels are MCP servers that implement the claude/channel capability. When you start Claude Code with the --channels flag, a polling service spins up alongside your session. Your phone sends a message → the app’s bot API delivers it → the local MCP plugin (a Bun process on your machine) wraps it as a <channel> event → Claude reads it, does the work, responds. Everything runs locally. Your code never leaves your machine.
Setup: Telegram in 5 minutes
Requirements: Claude Code v2.1.80+, Bun runtime, claude.ai login (API keys are not compatible).
# Install the plugin
/plugin install telegram@claude-plugins-official
# Configure with your bot token (create it via BotFather on Telegram)
/telegram:configure <YOUR_BOT_TOKEN>
# Restart with the channel active
claude --channels plugin:telegram@claude-plugins-official
Send a DM to your bot on Telegram — it responds with a linking code. Enter it in your Claude Code session and you’re connected. Discord setup is almost identical, via the Discord Developer Portal.
Test it first with Fakechat
Before connecting Telegram or Discord, Anthropic includes a demo called fakechat that spins up a local chat UI in your browser — no external accounts, no tokens. It’s the fastest way to understand the event flow before exposing your terminal to any external service.
/plugin install fakechat@claude-plugins-official
claude --channels plugin:fakechat@claude-plugins-official
A UI opens on localhost. You type a message, you see it arrive at your Claude Code session, you see Claude respond. Once you understand the mechanics, you move to a real platform.
Use cases for devs
Remote task delegation — You start a long refactor or migration, go to a meeting, check from your phone. Claude messages you when it finishes — or when it gets stuck on something.
CI failure triage — You forward build failures to your session via webhook. Claude inspects the logs with the full context of the project and either fixes the problem automatically or sends you a diagnosis. It’s not just a notification: it’s real triage.
Permission relay — Channel servers that declare the permission capability can relay tool approval prompts to your phone. Claude pauses, asks you remotely, and resumes when you respond. For completely unattended work there’s --dangerously-skip-permissions, but use it only in environments you control.
Multiple simultaneous channels — You can pass multiple plugins to --channels separated by spaces. Telegram for personal use, a webhook channel for CI — both active in the same session.
What’s still not there
This is a research preview with real limitations: during the preview, --channels only accepts plugins from Anthropic’s official allowlist (custom channels require --dangerously-load-development-channels). There’s still no persistent background mode — you need to keep a terminal session open, so tmux or screen are the current workaround. Telegram doesn’t have a message history API. Slack, WhatsApp, and iMessage are requested but not available yet.
The plugin architecture points to where this is going. Anthropic published a Channels reference spec for building custom integrations, and community-developed channels are explicitly part of the plan.
The bottom line
OpenClaw proved that devs want to stay connected to their agent without being glued to the terminal. Claude Code Channels delivers exactly that — with native MCP integration, real security controls (sender allowlists, admin management for Teams/Enterprise), and setup that doesn’t require buying dedicated hardware.
It’s a research preview. The flag syntax may change. But the workflow shift it enables — from synchronous sessions to an always-on agent you message like a teammate — is the direction the whole ecosystem is moving.
Have you already set up Channels in your workflow? Did you connect it to Telegram, Discord, or are you waiting for it to come out of preview? Let us know in the comments.