Claude Code Channels is Anthropic's new MCP-based system that connects Claude Code sessions to external messaging platforms, starting with Telegram and Discord. Shipped in v2.1.80 on March 19, 2026, with permission relay following in v2.1.81 the next day, Channels turns Claude Code from a terminal-only tool into a remotely accessible AI agent -- one you can message from your phone and get filesystem-level responses.

This matters because the architecture of AI agent access is shifting. The question is no longer "how smart is the model?" but "how do you reach it when you need it?" Anthropic just answered that question with a system that runs entirely on your local machine, exposes no inbound ports, and gives Claude Code full access to your files, git repos, and MCP servers -- all triggered by a Telegram message. The implications for developers, teams, and the broader AI agent ecosystem are significant. And so are the limitations.

What Is Claude Code Channels?

Claude Code Channels is an MCP-based plugin architecture that allows external events to push into a running Claude Code session. Rather than requiring a developer to sit at a terminal, Channels lets messages arrive from connected platforms -- currently Telegram and Discord -- and routes them into the active session as structured events. Claude processes the message with full access to the local filesystem, MCP servers, and git, then replies through the originating platform.

The architecture is worth understanding because it explains both the strengths and the constraints. Each channel plugin runs as a subprocess spawned by Claude Code, communicating over stdio. For Telegram and Discord, the plugin polls the platform's API at regular intervals. No inbound ports are opened. No webhook endpoints are exposed. The entire system runs locally on the developer's machine, which means zero attack surface from the network side -- but it also means the session must remain open for messages to flow.

Two official plugins ship out of the box: telegram@claude-plugins-official and discord@claude-plugins-official. A third, fakechat, serves as a localhost demo for developers building custom channels. Custom channel development requires the --dangerously-load-development-channels flag, signaling that Anthropic treats unofficial channels as an advanced, opt-in feature.

How Channels Works: Setup to First Message

The setup flow is straightforward but has specific requirements. First, install the plugin:

/plugin install telegram@claude-plugins-official

Configure the bot token (via BotFather for Telegram, or a Discord bot token), then restart Claude Code with the channel enabled:

claude --channels plugin:telegram@claude-plugins-official

On first connection, the system generates a security code. Send that code to the bot from your Telegram or Discord account to pair the session. After pairing, messages arrive as <channel source="telegram"> events inside the Claude Code session. Claude sees the message, processes it with full context, and sends the response back through the bot.

Several constraints apply. Channels requires a claude.ai login -- API keys are not supported. The Bun runtime is required for pre-built plugins. Team and Enterprise accounts must explicitly enable channelsEnabled in their organization settings. And critically, there is no daemon mode. The Claude Code session must remain open for messages to flow, which means most production deployments will require tmux, screen, or a macOS LaunchAgent to keep the session alive.

Permission Relay: The v2.1.81 Follow-Up

The day after Channels shipped, Anthropic released v2.1.81 with permission relay -- a feature that forwards tool approval prompts to the remote platform. This is the mechanism that makes Channels usable for real work rather than just chat.

Without permission relay, Claude Code would hit a tool approval prompt ("Allow file write to src/index.ts?") and stall, waiting for terminal input that never comes because the user is on their phone. With permission relay, that prompt appears in Telegram or Discord. The user approves or denies from their device, and execution continues.

This is clever engineering, but it introduces friction that breaks certain workflows. Any task requiring multiple tool approvals becomes a sequence of phone notifications demanding individual responses. For developers who run Claude Code in bypassPermissions mode with trusted workspaces, the permission relay is unnecessary overhead. For security-conscious teams, it is exactly the right tradeoff -- every action requires explicit human approval, even remotely.

The OpenClaw Elephant in the Room

It is impossible to discuss Channels without addressing OpenClaw. Peter Steinberger's open-source platform, which launched in November 2025 and has accumulated over 200,000 GitHub stars, pioneered the exact model Channels now implements: connect an AI coding agent to messaging platforms for remote access. OpenClaw currently supports 20+ platforms including Telegram, Discord, WhatsApp, Slack, iMessage, and more.

The history adds context. Anthropic sent Steinberger a cease-and-desist over his project's original name, "Clawd," which referenced Claude. OpenAI subsequently hired Steinberger. VentureBeat described Channels as "an OpenClaw killer." The competitive framing is unavoidable.

But the technical comparison reveals a more nuanced picture. OpenClaw supports 20+ platforms; Channels ships with two. OpenClaw runs as a persistent daemon; Channels requires an open session. OpenClaw has a mature plugin ecosystem; Channels is brand new. On the other hand, Channels has native integration with Claude Code's permission model, hooks engine, and 1M-token context window. It does not need to proxy through an API -- it is the runtime.

The real question is not which platform is better today. It is whether first-party integrations will catch up to open-source velocity, or whether the community will continue to build faster than Anthropic can ship. History suggests the latter, but Anthropic's advantage is depth of integration with its own model -- something third parties cannot replicate.

What This Means for AI Agent Orchestration

Channels represents a philosophical shift in how Anthropic positions Claude Code. The tool launched as a terminal agent for developers at their desks. With voice mode, /loop, and the hooks engine in earlier March releases, it became a programmable automation layer. Channels now makes it a remotely accessible agent that can respond to events from anywhere.

This trajectory -- from tool to platform to agent infrastructure -- mirrors what the broader AI industry is converging on. The value of a language model is no longer just its reasoning capability. It is the surface area of its integrations, the reliability of its execution environment, and the breadth of ways humans can interact with it. Channels is Anthropic's bet that first-party messaging integration, even if limited to two platforms at launch, is a necessary component of that surface area.

For teams building AI agent systems, the launch validates a pattern: messaging platforms as the primary interface for agent interaction. Not dashboards. Not IDEs. Not web apps. The chat thread -- where humans already spend their time -- becomes the command surface for AI agents that have full access to development infrastructure.

The Gaps That Matter

Three limitations will determine how quickly Channels gains adoption. First, the lack of daemon mode means production deployments require workarounds. Running a persistent Claude Code session inside tmux or a LaunchAgent is straightforward for experienced developers but adds operational overhead that OpenClaw's daemon architecture avoids entirely.

Second, two platforms is a narrow launch. Teams using Slack, Microsoft Teams, or WhatsApp as their primary communication tool have no official path. Custom channel development is possible but flagged as dangerous, which signals that Anthropic is not yet confident in the stability of the plugin API.

Third, the permission model -- while secure -- creates friction for autonomous workflows. The multi-agent patterns that developers are building with Claude Code often rely on agents operating without per-action approval. Permission relay makes remote use safe but slow for high-throughput automation.

These are not dealbreakers. They are the expected constraints of a v1 launch from a company that prioritizes safety over speed. The question is how fast Anthropic iterates.

What Builders Should Do Now

If you are already running Claude Code, Channels is worth testing today. The Telegram setup takes under five minutes if you have a bot token. Run it in a tmux session, pair your phone, and send a request that touches your codebase -- a test run, a git status, a file search. The experience of getting a structured response on your phone from an agent with full filesystem access is qualitatively different from chatting with a cloud-hosted model.

If you are building on OpenClaw or a similar orchestration layer, Channels is not a replacement -- it is a signal. Anthropic is moving toward first-party platform integration, which means the plugin APIs, permission models, and session management patterns in Channels will likely influence how Claude Code evolves. Understanding the architecture now gives you a head start on adapting when Channels inevitably expands its platform support.

For teams evaluating AI agent infrastructure, the right framing is not "Channels vs. OpenClaw." It is: what does your agent need to do, and which integration model -- first-party native, open-source hub-and-spoke, or both -- serves that need? The answer may be different for every team.

Frequently Asked Questions

What is Claude Code Channels?

Claude Code Channels is an MCP-based plugin system introduced in Claude Code v2.1.80 (March 19, 2026) that allows external messaging platforms to push events into a running Claude Code session. It ships with official Telegram and Discord plugins, enabling developers to interact with Claude Code remotely through chat apps while the agent retains full access to the local filesystem, git, and MCP servers.

How do I set up Claude Code Channels with Telegram?

Install the plugin with /plugin install telegram@claude-plugins-official, configure your Telegram bot token (obtained via BotFather), then launch Claude Code with claude --channels plugin:telegram@claude-plugins-official. A security code will appear in the terminal -- send it to your bot on Telegram to pair the session. Messages then flow bidirectionally between Telegram and the Claude Code session. Requires a claude.ai login and the Bun runtime.

Does Claude Code Channels work as a background daemon?

No. Claude Code Channels does not include a daemon mode. The Claude Code session must remain open for messages to be received and processed. Most developers use tmux, screen, or a macOS LaunchAgent to keep the session persistent. This is a key architectural difference from OpenClaw, which runs as a standalone daemon process.

How does Claude Code Channels compare to OpenClaw?

OpenClaw supports 20+ messaging platforms, runs as a persistent daemon, and has a mature plugin ecosystem with over 200,000 GitHub stars. Claude Code Channels currently supports 2 platforms (Telegram and Discord), requires an open session, and launched on March 19, 2026. Channels' advantage is native integration with Claude Code's permission model, hooks engine, and 1M-token context window -- it runs inside the Claude Code runtime rather than proxying through an API.

What is permission relay in Claude Code Channels?

Permission relay, introduced in Claude Code v2.1.81 (March 20, 2026), forwards tool approval prompts from Claude Code to the connected messaging platform. When Claude Code needs permission to execute an action (like writing a file), the approval request appears in Telegram or Discord. The user can approve or deny from their device, allowing remote supervision of agent actions without terminal access.

Sources: Claude Code Channels Documentation, Channels Reference, Claude Code Changelog, MacStories, Implicator.ai, The Decoder, Techzine Global, DEV Community, Telegram Plugin GitHub

Stay ahead of the AI curve -- bookmark nevo.systems for daily intelligence on AI agents, tools, and the technologies reshaping software development.