Claude Code Channels: the OpenClaw Killer? My verdict after testing it
Mike Codeur
Anthropic just launched Claude Code Channels on March 20, 2026. The promise: control Claude Code from Telegram or Discord, straight from your phone. VentureBeat called it an "OpenClaw Killer." After testing it extensively, here's my honest analysis.
How it works
Claude Code Channels uses an MCP server that pushes events into your session. Unlike regular MCP (where Claude calls a tool), here the external message arrives and drops into the session.
The complete flow:
- You send a message from Telegram on your phone
- The message hits the Telegram Bot API (in the cloud)
- On your machine, the MCP plugin (running on Bun) polls the Bot API
- The message drops into your Claude Code session
- Claude executes and the response goes back through the same path
Key point: everything runs locally. No port opened on your machine.
Telegram setup in 5 minutes
- Create a bot via BotFather on Telegram
- In Claude Code: install the official Anthropic plugin
- Configure with your bot token
- Launch with the --channels flag
- Pair by messaging the bot
For Discord, it's almost identical. The gotcha: enable Message Content Intent in the Privileged Gateway Intents.
What Channels does well
- Ultra fast setup: 5 minutes, no Docker, no reverse proxy
- Native security: user ID allowlist, pairing code, nothing exposed
- Extensible architecture: MCP plugin system allows more platforms
- File support: images and documents up to 50 MB
The limits nobody talks about
1. Your machine must stay on. Close your terminal, bot is dead.
2. Manual launch every time. The channels command every single time.
3. Zero persistent memory. No identity files, no daily notes, no long-term memory.
4. No scheduled tasks. No crons, no heartbeats. The agent does nothing on its own.
5. One channel at a time. Telegram OR Discord, not both.
6. Claude only. No model choice. Locked into Anthropic's ecosystem.
Anthropic shipped 3 features, not 1
| Feature | Date | One sentence |
|---|---|---|
| Remote Control | Feb 24 | A remote window into your terminal |
| Dispatch | Mar 17 | A walkie-talkie to your desktop assistant |
| Channels | Mar 20 | A Telegram/Discord chatbot connected to your terminal |
| OpenClaw | Nov 2025 | An autonomous agent that works while you sleep |
The comparison table
| Remote Control | Dispatch | Channels | OpenClaw | |
|---|---|---|---|---|
| Machine must run | YES | YES | YES | NO (VPS 24/7) |
| Persistent memory | No | No | No | YES |
| Scheduled tasks | No | Basic | No | YES |
| Autonomy | No | No | No | YES |
| Multi-channel | No | No | No | YES (10+) |
| Model agnostic | No | No | No | YES |
My verdict
Channels is a good start. The MCP architecture is clean. But even on a VPS, you'd just have a pipe. To make it a real assistant, you need to build everything around it: memory, crons, skills, resilience. That's exactly what OpenClaw spent months building.
It's a brick, not the house.
The real signal is that Anthropic validates the concept of personal agents accessible via messaging. 200K GitHub stars on OpenClaw, and now Anthropic shipping 3 features in one month to respond. The race for personal agents is on.
Full video with live demo: https://mkc.sh/claude-channels?utm_source=blog
Newsletter The Agentic Dev: https://mkc.sh/the-agentic-dev?utm_source=blog