Claude Code Channels + OpenClaw: The Always-On Agent Stack

Anthropic shipped Claude Code Channels this week. Version 2.1.80. Telegram and Discord bridges that push events directly into a running Claude Code session, making it bidirectional and always-on.

If you've been running OpenClaw with Claude Code agents in Discord, you already know what this is. We've been doing it.

What Channels Actually Are

Channels are MCP servers that act as event bridges. Instead of Claude Code sitting idle waiting for you to type in a terminal, it can now receive messages from Discord, Telegram, webhooks - anything that speaks the MCP protocol. The agent stays running. Events come in. The agent responds. No polling, no manual invocation.

The HN thread hit 323 points with 184 comments, and the reaction was mostly "finally." Developers have wanted always-on coding agents for months. The use cases are obvious: CI/CD alerts that trigger automatic fixes, Slack messages that spawn investigations, Discord commands that kick off deployments.

OpenClaw Was Already Doing This

Here's the thing - OpenClaw has been the canonical host for always-on Claude Code agents via Discord since before Channels existed. The architecture is different (OpenClaw uses its own event routing rather than Anthropic's MCP Channel protocol), but the outcome is identical: a Claude Code agent that's permanently available, listening for events, and acting on them without human invocation.

What Channels does is formalize this pattern inside Anthropic's own tooling. That's good for the community. It means the "always-on agent" concept isn't a hack or a workaround anymore - it's a first-class feature of Claude Code itself.

And OpenClaw is the infrastructure layer that makes running these agents practical. You don't want to keep a laptop open 24/7 with a Claude Code process running. You want a server that manages agent lifecycles, restarts on failure, routes events from multiple channels, and provides observability into what the agent is doing.

That's exactly what OpenClaw does. Channels is the protocol. OpenClaw is the platform.

The Always-On Agent Stack in Practice

Here's what a production setup looks like today:

Layer 1 - Event Sources: Discord server, Telegram bot, GitHub webhooks, Stripe payment events, monitoring alerts. Whatever generates events your agent should react to.

Layer 2 - Claude Code + Channels: The agent process itself, running with Channels MCP servers configured for each event source. Receives events, processes them, takes action.

Layer 3 - OpenClaw: The host that keeps Layer 2 alive. Process management, health checks, restart policies, log aggregation, multi-agent orchestration if you're running more than one agent.

Layer 4 - Agent Tools: MCP servers for your specific use case. AgentPay MCP for payments. Database connectors. API clients. File system access. Whatever your agent needs to actually do its job.

Without Layer 3, you've got a Claude Code process that dies when your SSH session closes. With OpenClaw, you've got an always-on agent that survives reboots, handles errors gracefully, and scales to multiple agents if you need them.

The Pi Coding Agent Question

One HN commenter called Anthropic out: Pi Coding Agent had multi-channel support and 700+ third-party packages months before Channels launched. That's fair criticism, and it's worth paying attention to.

Pi (pi.dev) built the always-on coding agent experience first. They have a package ecosystem. They have the multi-channel pattern. Anthropic is catching up with Channels, not inventing it.

But here's what matters for your architecture decision: Claude Code has the developer mindshare. GSD just crossed 36,000 stars. claude-hud hit 8,900 stars this week. learn-claude-code is at 34,000+. The Claude Code plugin ecosystem is the fastest-growing developer distribution channel on GitHub right now.

Pi might have shipped first. Claude Code has the ecosystem gravity. And OpenClaw is the infrastructure that makes the Claude Code version production-ready.

What to Build Next

If you're running Claude Code agents on OpenClaw today, Channels unlocks three immediate patterns:

Scheduled agents. Connect a cron-style event source that fires Channels events on a schedule. Your agent wakes up, does its job, goes back to listening. Database backups, report generation, dependency audits - anything you'd put in cron but want an LLM reasoning about.

Monitoring responders. Pipe your alerting system (PagerDuty, Grafana, Datadog) into a Channels MCP server. Agent receives the alert, investigates the issue, attempts a fix, and posts the result back to Discord. The 3 AM page gets handled before you wake up.

Multi-channel coordinators. One agent listening on Discord, Telegram, and GitHub webhooks simultaneously. A customer asks a question on Discord. The agent checks the GitHub repo for relevant issues, cross-references Telegram support threads, and responds with full context. All automated, all always-on, all running on OpenClaw.

The pieces are in place. Claude Code brings the intelligence. Channels brings the event bridge. OpenClaw brings the hosting. The always-on agent stack isn't theoretical anymore. Go build something.

This article was written with AI assistance. All technical claims, code, and architectural decisions were validated by the author.