# Channels Setup Guide

Channels let external chat apps push messages into a running
Claude Code session. Your agent reacts to messages as if you
typed them locally. This is Claude Code's answer to OpenClaw's
15+ native messaging integrations.

## Supported channels (March 2026)

| Channel | Plugin | Platform requirement |
|---------|--------|---------------------|
| Telegram | `@anthropic-ai/claude-code-telegram` | Any OS |
| Discord | `@anthropic-ai/claude-code-discord` | Any OS |
| iMessage | `@anthropic-ai/claude-code-imessage` | macOS only |

## Prerequisites

1. **Bun** (JavaScript runtime, required for channel plugins):
   ```bash
   curl -fsSL https://bun.sh/install | bash
   ```

2. **Full Disk Access** (macOS, required for iMessage):
   System Settings > Privacy & Security > Full Disk Access >
   Add your terminal app (Terminal.app or VS Code)

3. **Claude Code** v2.1.80+ (channels were introduced here)

## Install a channel plugin

```bash
# In a Claude Code session, run:
/install @anthropic-ai/claude-code-telegram
# or
/install @anthropic-ai/claude-code-discord
# or
/install @anthropic-ai/claude-code-imessage
```

Verify with `/plugins` to see installed plugins.

## Launch Claude Code in channel mode

```bash
claude --channels
```

You should see: `Listening for channel messages from plugin [channel-name]`

Claude Code now receives messages from the configured channel
and responds through it. Every tool, skill, and agent you have
configured works exactly as if you were typing locally.

## Permission approval from your phone

When Claude Code needs to run a sensitive action, it sends a
permission request through the channel. You can reply "yes" or
"no" from your phone to approve or deny.

To avoid constant prompts, launch with Auto Mode:
```bash
claude --channels --enable-auto-mode
```

## The critical limitation: session persistence

**If the Claude Code session closes, the channel stops working.**
Messages sent to the channel will not reach Claude Code until
you start a new session with `--channels`.

This is the biggest practical difference from OpenClaw, which
runs as a daemon and stays alive indefinitely.

### Workarounds for always-on

| Approach | How | Reliability |
|----------|-----|-------------|
| tmux/screen | `tmux new -s claude && claude --channels` | Good, survives SSH disconnect |
| Mac Mini | Dedicated machine, disable sleep | Very good |
| macOS VPS | Cloud Mac (e.g., MacStadium, AWS EC2 Mac) | Best, but expensive |
| /schedule | Remote triggers via claude.ai (no channel) | Different approach |

For iMessage specifically, you need macOS. A Linux VPS with
tmux works for Telegram and Discord but not iMessage.

## Channels vs Dispatch vs Remote Control

These three features all let you interact with Claude Code
remotely, but they work differently:

| Feature | Trigger model | Best for |
|---------|--------------|----------|
| **Channels** | Event-based: chat apps push messages into session | Reacting to messages, CI events, chat commands |
| **Dispatch** | Message-based: you send a task from mobile app | Delegating work while away |
| **Remote Control** | Direct control: you steer an active session | Guiding in-progress work from another device |

**Channels** = "my agent listens to a chat app"
**Dispatch** = "I send my agent a job"
**Remote Control** = "I take the wheel remotely"

## Enterprise and Team plans

Channel use must be enabled by an admin:
Organization Settings > Claude Code > Channels > Enable

By default, channels are disabled on Team and Enterprise plans.