Anthropic launched Claude Channels today—a research preview feature that pushes real-time events into Claude Code sessions from Telegram, Discord, or custom webhooks. Instead of Claude only responding when you type commands, it now reacts to CI builds, monitoring alerts, and chat messages while you’re away from your terminal. This turns Claude from a request-response tool into an event-driven autonomous agent. It’s trending #4 on Hacker News right now with 341 points, and it’s already sparking debates about security, cost, and whether always-on AI agents are sustainable.
What Claude Channels Actually Are
Claude Channels are MCP servers that push events into your Claude Code session, not tools that Claude calls. That architectural flip changes everything. Traditional tools sit idle until Claude invokes them. Channels actively monitor external systems—Telegram chats, GitHub webhooks, monitoring alerts—and forward events to Claude without human intervention.
Channels are also bidirectional. Claude doesn’t just receive events; it can reply back through the same channel. Send a question via Telegram, Claude analyzes your codebase, and the answer appears in your Telegram DM. The terminal shows the tool call (“sent”), but the actual reply only appears on the external platform.
This eliminates constant context rebuilding. Claude maintains session state across events instead of starting fresh every time you open a terminal. That’s the difference between an assistant you summon and an agent that operates continuously.
How to Set Up Claude Channels: Start with Fakechat
Before connecting real platforms, test the flow with Fakechat—a localhost demo channel that requires zero authentication. Install the plugin, start Claude with channels enabled, and type messages in a browser at http://localhost:8787. Claude receives the event, responds, and the reply appears back in the chat UI.
Here’s the complete setup:
# Install Bun (required for channel plugins)
curl -fsSL https://bun.sh/install | bash
# Install Fakechat plugin
/plugin install fakechat@claude-plugins-official
# Restart Claude with channel enabled
claude --channels plugin:fakechat@claude-plugins-official
# Open http://localhost:8787 and send a messageThe message arrives in your Claude session as a <channel source="fakechat"> event. Claude processes it, executes any work (file reads, analysis, edits), and calls the reply tool. The answer shows up in the browser, not the terminal. Once you’ve seen the flow work, graduate to Telegram or Discord for real integrations.
Real Setup: Telegram (The Preferred Choice)
Developers on Hacker News were surprised Anthropic chose Telegram over Slack, but the reasoning is sound. Telegram has 1 billion monthly users (vs Slack’s 50 million), and its bot API is trivial to work with. One commenter called it “the best programmatic integration” among chat platforms, and that shows in the setup process.
Create a bot via BotFather in Telegram. Send /newbot, give it a name and username ending in “bot”, and copy the token. Then configure Claude:
# Install Telegram plugin
/plugin install telegram@claude-plugins-official
# Configure bot token
/telegram:configure YOUR_BOT_TOKEN
# Restart Claude with channel enabled
claude --channels plugin:telegram@claude-plugins-officialDM your bot in Telegram. It replies with a pairing code. Back in Claude, approve the code and lock down access so only your account can send messages:
/telegram:access pair ABC123
/telegram:access policy allowlistNow only your Telegram ID can push events. Everyone else gets silently dropped. The security model is simple: pairing creates a sender allowlist, and Claude ignores any message from an unpaired ID.
Discord works similarly—create a bot in the Discord Developer Portal, enable Message Content Intent, set OAuth2 permissions (View Channels, Send Messages, Read History, Attach Files), and follow the same pairing flow. The Telegram plugin source and Discord plugin source are both on GitHub if you want to inspect the implementation.
Why This Matters: Real Use Cases
CI/CD Integration: Forward GitHub Action results to Claude via webhook. Test suite fails, Claude analyzes the errors, suggests fixes, commits changes, and triggers a re-run. Complete autonomous loops from code to test to fix to merge, with no manual context switches.
Chat Bridges: Your team DMs a Telegram bot with code questions. Claude analyzes the codebase and replies via the channel. Multiple people share a single Claude session without needing individual licenses or installations. Centralizes knowledge and keeps everyone in sync.
Monitoring and Alerting: Route Datadog or PagerDuty alerts to Claude. Production spike detected, webhook fires, Claude receives the event, analyzes logs, identifies root cause, and files a ticket. Faster incident response with full repository context instead of blind triage.
Event-Driven Workflows: Long-running refactoring tasks with approval checkpoints. Claude starts work, hits a decision point, sends an approval request via Telegram, waits for your reply, then continues or rolls back based on your response. True event-driven architecture without polling loops.
The Gotchas (And There Are Several)
Channels are a research preview. Features may change, and only Anthropic-maintained plugins are allowed during the preview period. Want to build a custom channel? Use --dangerously-load-development-channels for testing, but it won’t work in production until custom plugins are officially supported.
Known issues from the Hacker News thread: Remote Control websockets disconnect every 10 minutes, GitHub connector reports are “unavailable or non-functional,” and documentation has gaps. The economic model is unclear—will subscription plans cover always-on usage when this scales? Security teams are already asking hard questions about developers “hooking into personal machines via untrusted chat” and the lack of audit trails for what happened while you were away.
Enterprise adoption needs SOC 2 compliance docs, support for enterprise chat platforms (Slack Enterprise Grid, Microsoft Teams), and clearer usage analytics. Team and Enterprise organizations have channels disabled by default, and admins must explicitly enable the channelsEnabled setting before users can opt in with --channels.
Requirements are strict too: Claude Code v2.1.80 or later, Bun runtime installed, and claude.ai login (API keys aren’t supported). If you’re on Console or using API key authentication, this won’t work for you.
What Comes Next
Anthropic is clearly responding to competitive pressure from OpenClaw, which hit 210k GitHub stars after going viral in January. Multiple Hacker News commenters noted that Channels “mirrors OpenClaw’s architecture” and that “Claude caught up pretty quickly.” That’s not a coincidence. The AI agent ecosystem is converging on event-driven, persistent, autonomous architectures, and Channels is Anthropic’s entry into that race.
The Model Context Protocol foundation is solid—donated to the Linux Foundation in December 2025, used in production by Claude Desktop and Cursor IDE, and supported by 5,000+ community MCP servers. Channels extend that with event-push capabilities, and the research preview will likely graduate to production once the bugs get fixed and custom plugin support lands.
For now, start with the Fakechat demo to understand the flow. Graduate to Telegram if you want a real integration. Keep expectations realistic about stability and enterprise readiness. And watch the Hacker News discussion for community feedback—developers are already building webhook integrations, GitHub connectors, and multi-agent orchestration on top of this.
Key Takeaways
- Claude Channels launched today (March 20, 2026) as a research preview for pushing real-time events into AI sessions
- Setup is straightforward: Install Bun, add plugin, configure token, pair your account via allowlist
- Use cases are powerful: CI/CD automation, team chat bridges, monitoring integration, event-driven workflows
- Security model is simple: Pairing creates sender allowlist, only approved IDs can push events
- Research preview means bugs exist: Websocket disconnects, documentation gaps, economic model unclear
- Enterprise adoption needs work: SOC 2 docs, enterprise chat support, audit trails, admin controls
- Start with Fakechat demo: Test localhost flow before connecting real platforms (Telegram, Discord)
Claude Channels turn your AI assistant into an event-driven agent. It launched today, it’s powerful, it’s buggy, and it’s worth trying anyway.
