Claude Code Channels: how Anthropic is bringing its coding agent to Telegram and Discord

Home » Python » Claude Code Channels: how Anthropic is bringing its coding agent to Telegram and Discord

Anthropic is quietly reshaping how developers interact with its coding agent by rolling out Claude Code Channels, a system that wires active sessions into Telegram and Discord. Instead of being tied to a desktop terminal or IDE, you can now nudge an ongoing Claude Code session from your phone and get results back inside a familiar chat app.

This new capability effectively turns Claude Code into a remote, always-on coding companion that lives on your Mac or server but answers you through messaging bots. The feature is still in a research preview phase, and setup takes a bit of work, yet early reports suggest it already delivers a surprisingly capable, OpenClaw-style experience with more guardrails and official support.

What Claude Code Channels actually does

At its core, Claude Code Channels is a bridge between an existing Claude Code session and external chat platforms. You start Claude Code on your Mac (or another machine) and attach one or more channels; Telegram and Discord are supported out of the gate, with room for more services later on.

Once the link is established, anything you type into the paired Telegram or Discord bot is injected into the active Claude Code session as a channel event. Claude then executes commands, manipulates files, runs tools and sends the outcome back to you in the same chat thread, while also updating the terminal session on the host machine.

This makes it possible to treat Claude Code as a remote developer environment controller that you can ping from the couch, on the commute, or while traveling. As long as the underlying session is alive, the agent keeps listening for messages coming through the configured channel.

From Anthropic’s perspective, Channels brings Claude Code much closer to the “always-available AI worker in your pocket” experience that helped popularize open projects like OpenClaw, but implemented on top of Anthropic’s own agentic stack and security model.

How Channels works under the hood: MCP and Bun

Claude Code Channels is built on top of Anthropic’s Model Context Protocol (MCP), an open standard for connecting models, tools and data. MCP was originally introduced as a general-purpose way to let AI systems talk to external resources in a structured way, and was later donated to the Agentic AI Foundation under the Linux Foundation.

In the Channels design, an MCP server acts as a two-way gateway between the outside world and the running Claude Code session. Incoming chat messages are converted into MCP events and passed straight into the session; responses and task results are sent back out through specialized tools that know how to talk to Telegram or Discord.

Technically, the Channels feature relies on the Bun JavaScript runtime, which Anthropic acquired and now uses as the engine for the MCP servers involved. When you start Claude Code with Channels enabled, you are effectively launching a long-lived polling service that continuously watches for new messages from the configured plugins.

Because the MCP server and Bun process can run in a background terminal or on a persistent VPS, a Claude Code session can stay alive well beyond the lifetime of a typical web chat. Instead of timing out after inactivity, the agent just waits quietly until the next ping arrives from Telegram or Discord.

Version requirements and current preview status

Anthropic is positioning Channels as a research preview for now, available starting from Claude Code version v2.1.80. Users need a claude.ai account to sign in, and the feature is disabled by default for enterprise and team plans unless administrators explicitly turn it on in the admin console.

This cautious rollout reflects the nature of the feature: Channels gives a chat app the ability to trigger real actions in a developer environment, from executing CLI tools to modifying project files. Anthropic appears to be leaning on MCP’s structure and permission systems, along with whitelisting, to reduce the risk of misuse.

Despite the “preview” label, hands-on testing suggests that the integration is already fairly robust. The trade-off is that setup still feels a bit technical: users need to install plugins, configure bots and ensure Bun is properly installed before everything starts working.

Setting up Claude Code Channels with Telegram

On Telegram, the process begins with creating a bot and connecting it to Claude Code through an official plugin. For anyone who has dealt with Telegram bots before, the initial steps will feel familiar:

• Open the @BotFather account on Telegram.
• Use the /newbot command.
• Choose a bot username that ends with “bot” and set a display name.
• Copy the API token that BotFather generates.

On the machine where Claude Code runs, you then install the official Telegram plugin and reload your plugins so that Claude can see the new connector. The next step is to paste the API key from Telegram into the plugin configuration and complete a pairing process, which usually involves sending a direct message to your bot to obtain a one-time code.

That pairing step is important: Anthropic uses a whitelist mechanism to associate specific Telegram user IDs with the Claude Code session. Only approved accounts can send messages that Claude will actually process; other users, even in the same group chat or server, are silently ignored.

Once everything is wired up, you can start Claude Code with a command that includes the channels flag, for example:

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

From that point on, messages you send to the bot flow directly into your active terminal session. Responses show up both in the terminal and inside Telegram, keeping both interfaces in sync.

Discord integration and the MCP plugin model

The Discord setup follows the same basic pattern, but adds a few platform-specific steps in the Discord Developer Portal. You need to register an application, create a bot user, configure the necessary permissions and invite the bot to a specific server.

After that, you install the official Discord plugin from Anthropic’s repositories in essentially the same way as the Telegram one and add it to the Channels configuration when launching Claude Code. As with Telegram, a pairing sequence ties your personal Discord account to the channel, ensuring Claude only responds to authorized senders.

The interesting aspect here is that Telegram and Discord plugins are just MCP-based connectors hosted by Anthropic on GitHub. Because they follow the MCP standard, the broader community can build additional connectors for other platforms such as Slack or WhatsApp using the same pattern.

This makes Channels feel less like a one-off feature and more like the first wave of a broader ecosystem of messaging integrations. The MCP layer gives Anthropic and third-party developers a consistent way to plug new channels into Claude Code without redesigning the underlying agent every time.

What it’s like to use Claude Code via Telegram

Once Channels is configured, actually working with Claude Code through Telegram or Discord feels very similar to chatting with a friend who happens to control a powerful dev machine. You send a message, the bot shows a “typing…” indicator, and a few moments later, you get a detailed response or update.

Telegram supports attachment uploads up to 50 MB, including images that Claude can view inline. Long messages are automatically split into multiple chunks so that the conversation remains readable. If a task will take some time, Claude can send an interim “Processing…” message and later edit that same message to include the final result instead of flooding the chat with separate updates.

In practical tests, developers have used a Telegram-connected Claude Code session to run complex workflows entirely from a phone. Examples include building and running an iOS project with xcodebuild and deploying it wirelessly to the very device used to chat, querying saved articles via the Readwise Reader CLI, and kicking off podcast transcription skills that return cleaned transcripts, subtitle files and markdown reports back into Telegram.

Because the heavy lifting still happens on the Mac or server, Channels gives iOS and Android users access to workflows that would otherwise be impossible on mobile alone. Debugging an iPhone app by sending a screenshot to Claude in Telegram, for instance, becomes practical even when you’re away from your development machine.

There are limits: voice messages are not yet supported, and if Claude Code prompts for input directly in the terminal, you may still need to interact with the host machine. But for many day‑to‑day operations, the chat interface is enough to keep work moving.

Permissions, safety flags and whitelists

Channels introduces a subtle but important tension between convenience and safety. When Claude Code needs explicit permission to perform certain file or system operations, there is currently no way to grant those permissions purely through Telegram or Discord.

In practice, this means you sometimes have to return to the host terminal to confirm an action before the agent can proceed. For power users already comfortable with Claude Code’s risk profile, one workaround is to start sessions with a flag like --dangerously-skip-permissions, which bypasses many prompts and lets Claude act more autonomously.

Using that flag does increase risk, particularly when combined with a messaging interface, so Anthropic emphasizes other layers of protection such as per-channel sender whitelists. Each channel keeps track of which specific user IDs are allowed to issue commands, rather than granting blanket access to a group or server.

Even within the same Discord server or Telegram chat, messages from unpaired members are quietly dropped and never reach the Claude Code session. This design is meant to prevent accidental or malicious triggers from other people in shared spaces, even if they can see the bot.

For organizations, the default requirement for admins to enable Channels on team or enterprise accounts acts as another safeguard. Companies that want this capability can turn it on and set policies, while others can leave it disabled until they are comfortable with the implications.

How Channels compares to Remote Control and terminal emulators

Before Channels, developers already had a few options for remote interaction with Claude Code. Terminal emulators, SSH sessions and the Remote Control feature in the Claude mobile app let users access a Mac’s terminal or desktop from a phone, albeit with some friction and, in some cases, stability concerns.

These approaches tend to offer more direct visibility into what is happening in the terminal window, and they are better suited for situations where you must respond to detailed permission prompts or inspect log output in real time. However, they are often less convenient for quick, conversational requests.

Channels, by contrast, is aimed squarely at people who want to interact with Claude Code in the same casual way they text a colleague. You do not have to juggle remote desktop apps, maintain a constant connection or worry about window focus. You just send a message and wait for Claude to report back.

That said, Channels does not replace traditional remote tools entirely. For complex debugging sessions, low-level system configuration or sensitive operations that you prefer to supervise closely, a full remote terminal or desktop remains more appropriate.

The broader trend is that developers now have a spectrum of ways to work with AI agents, from chat-based Channels to classical terminals, and can choose the right one based on the task and their risk tolerance.

OpenClaw, Dispatch and the “lobster-style” agent trend

The arrival of Claude Code Channels is widely seen as Anthropic’s answer to OpenClaw and similar “lobster-style” AI agents that run continuously in the background and interact through messaging apps. OpenClaw, created by Austrian developer Peter Steinberger, gained attention in late 2025 for letting people direct a personal AI worker via iMessage, Slack, Telegram, WhatsApp and Discord.

OpenClaw and related tools did more than just chat. They executed real work autonomously: drafting and sorting emails, managing files, building full applications, submitting job applications on behalf of users and orchestrating complex marketing campaigns across social networks, then notifying users through their messenger of choice as soon as each job finished.

That flexibility came at a cost. Observers flagged significant security issues, including broad file system access and relatively complex installation and configuration steps. Community efforts like NanoClaw emerged to tame some of those risks, but the basic pattern—an always-on agent steered from chat—remained attractive to many early adopters.

Anthropic has been gradually assembling its own answer. In the span of a week, the company rolled out Dispatch, Channels, a new voice mode, a /loop feature, expanded 1M-context support, MCP guidance, persistent Cowork sessions on mobile, cross-application context for Office tools and inline charting. Taken together, these updates push Claude beyond a traditional chat interface toward a more agentic, task‑oriented platform and help teams automate code reviews.

Commentators have noted that Dispatch alone already satisfies a large portion of what users liked about OpenClaw. By adding Channels on top, Anthropic is effectively offering the same around-the-clock, messenger-driven control pattern, but wrapped in an enterprise-ready framework and backed by a commercial provider.

Always-on agents and the limits of today’s Channels

Despite the similarities, there is an important difference between Claude’s current approach and some “lobster-style” systems. OpenClaw uses a local daemon process to keep the agent alive continuously, even when a user is not actively interacting with it. In contrast, both Dispatch and Channels still depend on the underlying Claude Code session remaining active.

In practical terms, if the terminal window closes, the machine goes to sleep or the server shuts down, the channel connection is lost and messages sent from Telegram or Discord will no longer reach Claude. Users can mitigate this by running Claude Code on a persistent VPS or a Mac configured to stay awake, but true fire-and-forget operation is not fully there yet.

Industry watchers see this as a waypoint on a broader trajectory toward more autonomous, cloud-hosted agents. Other players have already been exploring similar territory: OpenAI’s Codex products have supported cloud-based asynchronous tasks, and Google’s Jules project has run agents inside virtual machines that continue processing after the user steps away.

OpenClaw helped normalize the idea that you might simply message an AI and have it perform work in the background, reporting back only when necessary. Anthropic’s Channels appears to be moving in the same direction, but with more emphasis on clear standards (through MCP) and governance suitable for commercial and enterprise use.

Once users grow comfortable with the pattern of “send a message, have someone else do the work”, simple static chat windows may start to feel inadequate. The competitive focus shifts from model quality alone to orchestration, persistence and the ability to integrate with everyday communication tools.

Developer experience: from desks to pockets

For everyday developers, Channels translates into a subtle but meaningful change in how and where coding-related work happens. Many tasks that used to require being physically in front of a Mac—kicking off builds, running test suites, tidying files, checking logs—can now be triggered from the phone in your pocket.

Some users describe it as a kind of “superpower” for iOS or Android: you can leverage tools and command-line utilities that mobile platforms do not natively provide, while still staying within a comfortable messaging interface. The host machine quietly carries out the heavy lifting and reports back in discrete updates.

This is also one reason why Apple’s lack of a mobile Xcode offering has drawn renewed attention. As coding on Apple platforms continues to depend heavily on macOS tooling, remote agents like Claude Code Channels fill the gap for developers who want to stay productive while away from their main workstation.

At the same time, Channels blurs the line between “coding” and “coordination”. Developers can assign tasks, review summaries or request quick code changes in the same apps where they already discuss product decisions or track issues with teammates. Over time, this may nudge workflows toward a more conversational style of development.

Given Claude Code’s existing popularity as a command-line and IDE companion, the addition of official, supported messaging connectors is likely to accelerate adoption. While the open agent ecosystem around OpenClaw remains vibrant, its user base is fractured across many variants, and some observers argue that a polished, enterprise-leaning option like Channels could become the default choice for teams that prioritize reliability and governance.

All of this leaves Anthropic in a position where Claude Code is no longer just a tool you open at your desk, but a partner you can reach from practically anywhere. With Channels, dispatch-style task scheduling and a growing stack of MCP integrations, the company is betting that the future of AI-assisted development will look less like filling in prompts in a browser and more like messaging a colleague who quietly handles the grunt work in the background.