Claude Code Channels: Control Your AI Agent via Discord and Telegram

Anthropic just shipped a feature that fundamentally changes how marketing teams interact with AI agents: Claude Code Channels, a system that lets you push messages, alerts, and webhooks directly into a running Claude Code session from Discord, Telegram, or any custom HTTP endpoint. As VentureBeat reported on March 20, 2026, Anthropic’s move is positioned as a direct answer to the popular open-source autonomous agent OpenClaw — and for practitioners who have been waiting for a reliable, authenticated way to command AI agents from their phones or team chat, this is the unlock.

What Happened

On March 20, 2026, Anthropic shipped Claude Code Channels as a research preview feature inside Claude Code v2.1.80 and later. The announcement, noted by VentureBeat, frames Channels as a direct competitive response to OpenClaw — the open-source autonomous agent that had been popular precisely because it let developers interact with a running AI agent from messaging platforms. Anthropic’s version arrives with tighter security, an official plugin system, and deep integration with Claude’s existing toolchain.

Here is exactly what Channels does, stripped of marketing language: a Channel is an MCP (Model Context Protocol) server that runs on the same machine as your Claude Code session and acts as a bridge between the outside world and Claude’s working context. You configure the bridge — say, a Telegram bot or a Discord bot — and Claude Code then reacts to incoming messages as if they were prompts typed directly at the terminal. The session keeps running in the background, and whenever someone (an allowlisted human or an automated system) pushes a message into the channel, Claude picks it up and acts on it.

Two modes are supported per the official Channels documentation, and the distinction matters for how you design workflows:

One-way channels forward events into Claude’s session without expecting a reply. The primary use cases are CI/CD alerts, monitoring webhooks, and scheduled job results. Your analytics platform fires an anomaly alert, it arrives in Claude’s session as a structured <channel> event, Claude reads it, takes action (runs a diagnostic, creates a task, posts a summary), and you see the result in your terminal. No human needed to initiate the interaction.

Two-way channels are full chat bridges. Claude reads the inbound message and can reply back to the same platform using a registered MCP tool. The officially supported two-way channels at launch are Telegram and Discord. Both are open-source and available in the claude-plugins-official GitHub repository.

The technical architecture is notably practical. Per the Channels documentation, chat-based channels (Telegram, Discord) run as local processes on your machine and poll the platform’s API for new messages. Nothing is exposed to the public internet — there is no inbound port opened on your machine, no webhook endpoint that needs to be publicly reachable. Webhook-based channels listen on a localhost HTTP port, so your external systems POST to 127.0.0.1:PORT via an internal tunnel. Anthropic provides a demo called “fakechat” — a localhost web UI — that lets you test the entire channel system before touching any real messaging platform.

Setup for Discord, per the official setup steps, follows a clear sequence: create a Discord bot in the Developer Portal, enable Message Content Intent, invite the bot to your server with the correct permissions (View Channels, Send Messages, Send Messages in Threads, Read Message History, Attach Files, Add Reactions), install the plugin inside Claude Code with /plugin install discord@claude-plugins-official, configure your bot token with /discord:configure <token>, and restart Claude Code with the --channels flag pointing to the Discord plugin. A pairing code system then locks the bot to specific Discord user IDs — anyone else in the same server who tries to message your bot gets silently dropped. Telegram setup is nearly identical: create a bot through BotFather, install the plugin, configure the token, pair your account.

One important constraint noted in the Channels research preview section: Channels currently only accepts plugins from an Anthropic-maintained allowlist. During the research preview, custom channels require a --dangerously-load-development-channels flag, and they go through security review before being added to the official marketplace. This is deliberate — as the Channels reference documentation states explicitly, “an ungated channel is a prompt injection vector.” The sender allowlist architecture is the core security guarantee, and Anthropic is treating prompt injection as a first-class threat model rather than an afterthought.

Why This Matters

The problem Channels solves is not a developer problem — it is an operational problem, and marketing teams have been living inside it since AI agents became usable. AI agents are only as valuable as your ability to task them and receive their output. If the only way to interact with your agent is to sit at a terminal, the agent is effectively desk-bound. Every time you step away — into a client call, a strategy meeting, a cross-country flight — your agent stops being useful.

Channels changes the access model. A running Claude Code session becomes something you can command from wherever you already have a messaging app open. For marketing teams specifically, the implications break down across three dimensions:

Operational mobility. A growth marketing team managing paid search, email sequences, content pipelines, and social scheduling cannot afford to be terminal-bound. With a Telegram channel wired to their Claude Code instance, a performance marketer on mobile can fire a prompt like “pull yesterday’s Google Ads data, summarize by campaign, and flag anything with a CPA spike above 20% week-over-week” while waiting at an airport gate. Claude executes locally, with full access to the codebase, MCP tools, and any data connectors already configured in the session. The result comes back in Telegram. The analyst never needed to open a laptop.

Team and agency workflows. Discord is the de facto communication layer for many digital agencies and remote marketing teams. Claude Code Channels means you can build a bot that lives in a dedicated #ai-tasks channel, accepts briefs from team members (gated by Discord user ID allowlists), executes Claude Code sessions against your real project files and data, and posts results back in the thread. This is a categorically different AI integration than a chatbot — it is an authenticated agent with full code execution, file access, and connected tools, accessible through the interface your team already uses for everything else.

Webhook-driven pipeline automation. One-way channels are where the marketing automation story gets serious. Marketing stacks generate enormous event data: campaign completion notifications from your email platform, anomaly detections from your analytics layer, content approval signals from your CMS, A/B test completion events from your experimentation tool. All of those can become Claude triggers. A webhook from your email platform fires when a campaign hits a 5% unsubscribe rate — Claude receives it, analyzes the subject line and send-time cohort data, drafts a root-cause summary, and posts it to your Slack via an MCP server. No human has to notice the alert, interpret it, or route it. The analysis is already done when the human sees the notification.

The enterprise access controls reveal Anthropic’s seriousness about production deployment. Per the enterprise controls documentation, for Pro and Max individual plans, Channels is opt-in per session using the --channels flag. For Team and Enterprise plans, it is disabled by default until an administrator enables it through claude.ai admin settings — a governance model that lets IT and security teams control the blast radius before rolling out to the broader organization. Admins toggle it from claude.ai → Admin settings → Claude Code → Channels, or by setting channelsEnabled: true in managed settings.

What assumptions does this challenge? The core one: that AI agents are tools you go to, rather than tools that work for you asynchronously and meet you in the platforms you already live in. Every significant AI-native workflow that marketing teams have built to date requires a human to initiate the session, wait for output, and manually act on it. The Channels model is closer to having a capable analyst on call — one who stays engaged between check-ins and responds when you reach out, from wherever you are.

The Data

The following table maps Claude Code’s remote access and integration surfaces against each other, based on the official Claude Code documentation. This comparison is directly relevant for marketing operations teams evaluating which integration pattern fits which workflow.

Sources: Channels docs, Remote Control docs, Slack integration docs, Scheduled tasks docs

The table makes the positioning clear. Remote Control is the right choice when you want to continue local work from another device with full interactive control. Slack is best for team-facing task delegation tied to GitHub repos. Scheduled tasks handle polling and time-based automation. Channels is the one to reach for when external events — from messaging apps or external systems — need to drive an always-on agent without human initiation at the moment of tasking.

A second data point worth noting from the Channels reference documentation: the notification event format that Claude receives contains a content field (the message body) and a meta field (arbitrary key-value attributes like chat_id, severity, or sender). This structured event format means Claude gets context beyond just the raw message — it knows which conversation to reply to, what severity level the alert carries, or any other routing metadata your channel server includes. This matters for building reliable workflows where Claude’s response needs to be precisely targeted.

Real-World Use Cases

Use Case 1: Content Pipeline Management via Telegram

Scenario: A solo content marketer runs a newsletter, a blog, and a LinkedIn presence. They use Claude Code locally for drafting, editing, and scheduling — but their best ideas hit them away from the desk. The current workflow means ideas die in voice memos or get forgotten during the commute.

Implementation: They set up a Telegram bot connected to a Claude Code session that runs persistently via a background terminal on their home machine. The Claude session has access to their content calendar (a local markdown file managed via a custom MCP tool), their publishing scripts (Ghost CMS API, LinkedIn via another MCP tool), and their draft folder with all historical posts. When an idea strikes during a commute, they Telegram the bot: “Add to draft queue: ‘Why AI agents need human-in-loop review for regulated industries’ — target publish next Thursday, 1200 words, practitioner tone.” Claude receives the message, creates a draft file in their drafts folder with the brief pre-filled, updates the content calendar markdown, and replies in Telegram with a confirmation and the scheduled publish slot.

Expected Outcome: Idea capture rate improves substantially because the friction of opening a laptop, navigating to the project, and starting a Claude session is eliminated. The content calendar stays current without manual entry. Over a month, the marketer estimates recovering two to three publishable ideas per week that would otherwise have been lost.

Use Case 2: Agency Performance Alert Triage via Discord

Scenario: A performance marketing agency runs paid campaigns for 15 clients across Google, Meta, and TikTok. The team communicates in Discord. They want Claude to handle first-pass triage when their analytics layer fires campaign anomaly alerts, before a human strategist reviews.

Implementation: The agency’s analytics platform (a custom Python reporting stack) posts anomaly alerts to a local webhook receiver running as a Claude Code channel server. When a campaign’s CPA rises more than 25% week-over-week, the webhook fires a structured payload containing the campaign ID, the metric, and the delta. Claude receives the event, queries campaign data via an MCP server connected to their BigQuery warehouse, generates a structured triage report (campaign name, affected ad sets, suspected root cause, recommended action), and posts it to the team’s #campaign-alerts Discord channel via a two-way channel reply. The senior strategist on call receives a fully-formed analysis in Discord rather than a raw alert number.

Expected Outcome: Mean time to triage drops from 45 minutes (waiting for a human to notice the alert, pull data, and analyze) to under 5 minutes for a complete first-pass analysis. The strategist’s job shifts from data retrieval to decision-making. Agencies managing high-volume campaigns can realistically monitor 3-4x more accounts with the same headcount.

Use Case 3: SEO Content Audits On Demand via Telegram

Scenario: An in-house SEO lead wants to run content audits on-demand from their phone — checking freshness, internal linking gaps, and keyword coverage — without waiting to get back to their laptop where their audit tooling lives.

Implementation: They configure a Telegram bot linked to a Claude Code session that has their static site repository locally cloned, an MCP server connected to their Google Search Console data, and a set of custom Claude Code skills (/audit-post, /check-internal-links, /keyword-gap). From Telegram, they send: “Run a content audit on /blog/ai-marketing-tools — check keyword freshness and internal links.” Claude runs the relevant skills, parses the GSC data for the URL, checks internal linking structure in the local repo, and replies in Telegram with a structured markdown summary of findings and a prioritized fix list.

Expected Outcome: Audit turnaround goes from “add to backlog, do it when at desk” to same-session completion regardless of location. Over a quarter, the SEO lead runs approximately 40% more audits than previously possible, catching content decay earlier and shortening the cycle from insight to editorial action.

Use Case 4: Email Campaign QA Before Launch via Discord

Scenario: An email marketing manager at an e-commerce brand runs 3-5 campaigns per week. Their final QA checklist — checking personalization tokens, unsubscribe links, preheader text length, subject line spam triggers, brand voice compliance — is manual and error-prone at volume.

Implementation: The email ops team sets up a Claude Code session with access to their Klaviyo API via MCP, their brand voice guidelines stored in a CLAUDE.md file, and a /qa-email skill that encodes the full checklist. When a campaign is ready for review, any team member DMs the Claude Discord bot with the campaign ID: “QA campaign ID 84721 before 3pm launch.” Claude calls the Klaviyo MCP tool to pull the campaign template, runs all checklist items systematically, flags issues with severity ratings (critical / warning / info), and replies in Discord with a structured QA report. If everything checks out, it replies: “Campaign 84721 passed all QA checks. Ready to launch.”

Expected Outcome: Manual QA time drops from 25 minutes per campaign to 3-4 minutes of human review of the Claude-generated report. Error rate on launched campaigns drops because the systematic checklist runs consistently rather than variably by whoever is doing QA that day. The bot’s Discord presence also makes QA status visible to the whole team without a separate status update step.

Use Case 5: Competitive Intelligence via Monitoring Webhooks

Scenario: A B2B SaaS company’s marketing team wants to monitor competitor activity — pricing page changes, new product announcements, new blog posts — and surface structured intelligence briefs in team Discord without requiring anyone to manually check competitor sites or set up calendar reminders to do so.

Implementation: The team runs a lightweight monitoring script (a simple cron job using a site-change detection library) that fires a webhook POST to their Claude Code channel server when a monitored competitor URL changes. The channel event contains the URL, a brief diff summary, and a timestamp. Claude receives the event, fetches the current version of the changed page via a web-fetch MCP tool, compares it to the cached previous version stored locally, writes a structured competitive intelligence brief (what changed, likely strategic implication, recommended marketing response), and posts it to the team’s #competitive-intel Discord channel.

Expected Outcome: Competitive response time improves from “discovered days later during a weekly review” to “alerted and analyzed within minutes of the change going live.” The marketing team can react to competitor moves — especially pricing changes and feature announcements — with enough lead time to adjust campaigns, update battle cards, and brief sales before customers start asking questions.

The Bigger Picture

Claude Code Channels lands at a specific moment in the AI marketing stack evolution. For the past two years, the dominant AI integration pattern in marketing has been the synchronous chat interface: open a browser tab, start a conversation, get output, copy it somewhere useful. Even the most sophisticated implementations of this pattern — custom GPTs, Claude Projects, embedded AI writing assistants — require the human to initiate every interaction and manually route the output somewhere actionable.

The next layer of the stack — the one that actually delivers on the “always-on AI agent” promise — requires agents that persist, react to external events, and integrate bidirectionally with the tools and communication platforms teams already use. Slack, Discord, and Telegram are not niche platforms; they are the operational nervous system of most digital marketing teams and agencies. An AI agent that lives inside those platforms, responds when messaged, and can also be triggered by automated events without human initiation is categorically more useful than one you have to navigate to.

This positions Channels alongside a broader set of moves Anthropic has been making with Claude Code in early 2026. The Agent SDK enables custom orchestration layers. GitHub Actions integration wires Claude into CI/CD pipelines. Remote Control allows session continuity across devices. Slack integration delegates coding tasks from team chat. Channels is the event-driven piece of this architecture — the ability for external systems to push work into a running Claude session without requiring a human to initiate it at the moment of tasking.

The comparison to OpenClaw, flagged by VentureBeat, is worth taking seriously. OpenClaw gained traction precisely because open-source autonomous agents demonstrated that the “command your AI from a messaging app” interaction model was viable and valuable — people genuinely wanted to work this way. Anthropic’s answer is to build the same capability into their first-party product with authentication, allowlist-based security, enterprise governance controls, and an extensible plugin architecture. The explicit prompt injection threat modeling in the Channels reference documentation signals that Anthropic is designing for production deployment, not just developer experimentation.

For marketing teams evaluating their AI stack, the practical implication is that the platform bet is clarifying. Anthropic is building a complete surface area for Claude Code: terminal, IDE, desktop app, web, mobile, Slack, GitHub, and now Discord/Telegram/webhooks. Teams that invest in Claude Code as their agentic infrastructure gain access to all of these surfaces as they ship, rather than maintaining separate integrations across different point tools. The Channels documentation explicitly maps each surface to its best use case — an architecture decision worth studying before building any new AI tooling.

The research preview status is honest signal. Anthropic has stated explicitly that the --channels flag syntax and the notification protocol contract may change based on feedback. For practitioners evaluating production deployment: the capability is real and working today on Claude Code v2.1.80+, but plan for API changes in Q2-Q3 2026 as the preview matures into a stable release.

What Smart Marketers Should Do Now

1. Map your team’s real async AI friction points before building anything.

Before configuring a Discord or Telegram channel, spend one week logging every moment where you or your team wanted to use AI but didn’t — because you were away from your laptop, in a meeting, or on mobile. These are your actual Channel use cases. The best channel setups solve documented friction, not hypothetical workflows. Build a short list of five specific interactions that currently require you to be at a terminal, then prioritize them by frequency and time cost. Design your initial channel architecture around the top two.

2. Start with a one-way webhook channel connected to your highest-signal alert.

The lowest-friction path to production value from Channels is a one-way webhook. Pick your most critical monitoring alert — campaign CPA anomaly, content publish failure, A/B test completion, server error spike — and wire it to a Claude Code channel server using the webhook receiver pattern documented in the Channels reference. This requires no chat platform setup, no bot account creation, and delivers immediate value by turning raw alerts into analyzed summaries. Build your team’s intuition for event-driven agent workflows before tackling the more complex two-way chat integration.

3. Build your Claude session’s context layer before you open the channel.

The quality of output from any Claude Code channel is entirely determined by what Claude has access to in the session — the MCP servers, the CLAUDE.md instructions, the local data files, the connected APIs. A Telegram message saying “summarize this week’s campaign performance” delivers dramatically different results in a session with a Google Ads MCP server, a BigQuery connector, and a detailed brand voice CLAUDE.md than it does in a vanilla Claude Code session with no context. Invest the time in the session infrastructure before you advertise the channel to your team. The channel is just the delivery mechanism; the session context is the actual product.

4. Design and document your sender allowlist policy before inviting anyone.

The Channels security model is explicit: only sender IDs you’ve paired can push messages into your session, and everyone else is silently dropped. If you’re setting up a team Discord channel, decide in advance which team members get access, pair each of them individually, and document the policy in writing. For agencies considering any form of client access to a Claude Code channel, be particularly deliberate about scope — a client who can message your Claude bot can, in principle, ask it to run any task the session has permissions for. Scope your session permissions tightly and pair only the client accounts you specifically intend to authorize.

5. Get your Team or Enterprise admin to enable Channels in admin settings today.

Team and Enterprise plans have Channels disabled by default. Per the enterprise controls documentation, the toggle location is: claude.ai → Admin settings → Claude Code → Channels. If you wait until you have a full use case designed and ready before requesting admin access, you’ll hit a setup blocker at exactly the wrong moment — when your team is ready to ship and you’re waiting for IT approval. Get the toggle flipped proactively so the capability is available when your team is ready to experiment with it.

What to Watch Next

Custom channel marketplace expansion. The current research preview restricts Channels to plugins from Anthropic’s official allowlist. Custom channels require --dangerously-load-development-channels and go through a security review to get listed. Watch the claude-plugins-official GitHub repository for new additions beyond Telegram and Discord. Likely next additions based on team communication trends: WhatsApp Business, Microsoft Teams, and SMS gateway plugins. The Channels reference documentation provides the complete MCP server contract for anyone who wants to build and submit a custom channel ahead of formal marketplace expansion.

API stability post-research-preview. Anthropic has explicitly flagged that the --channels flag syntax and the notification protocol contract may change. If you build internal tooling on top of Channels during the preview, maintain a thin abstraction layer between your code and the channel startup command. Watch the Claude Code GitHub releases for v2.2.x announcements, which will likely include the stable Channels API. Subscribe to the GitHub repository’s release notifications rather than relying on blog announcements.

Enterprise governance feature expansion. The current enterprise controls are binary: Channels on or off at the organization level. Over the next two quarters, expect more granular controls — per-channel approval policies (allow Discord but not Telegram), user-level channel access within an org, and audit logging for channel events. These are the features that unlock Channels for regulated industries (financial services, healthcare, legal, insurance) where every AI interaction needs a documented paper trail. If your org operates in a regulated vertical, this is the set of features to watch before committing to a production Channels deployment.

Channels and Scheduled Tasks convergence. Right now, Channels and Scheduled Tasks are documented and implemented separately, but the most powerful marketing workflows combine them: a scheduled task that runs every morning and posts its results back through a Discord channel, or a Telegram message that triggers a one-time scheduled reminder. The Claude Code team’s GitHub issues suggest tighter integration between these two features is on the roadmap. When it ships, the event-driven plus time-driven combination will cover essentially every async marketing workflow pattern without requiring multiple separate session management strategies.

OpenClaw ecosystem response. The open-source autonomous agent community is not standing still. OpenClaw and similar projects will respond to Channels with their own enhancements — likely focusing on areas where Anthropic’s first-party offering currently has gaps: broader messaging platform support out of the box, more flexible session persistence that survives machine restarts, and compatibility with non-Claude models for teams with multi-model stacks. For practitioners who need platform independence or open-source flexibility, watch the OpenClaw repository for its next major release. The competitive dynamic benefits the whole market by accelerating the development of event-driven agent architecture across the ecosystem.

Bottom Line

Claude Code Channels is the most meaningful step Anthropic has taken toward making Claude Code an operational tool for marketing teams rather than a developer-only productivity utility. By enabling bidirectional communication through Discord, Telegram, and custom webhook endpoints, Channels lets practitioners command AI agents from wherever their work actually happens — their phone, their team chat, their monitoring stack — without requiring anyone to sit at a terminal waiting for the next task to arrive. The security model is well-designed with sender allowlists and pairing codes. The plugin architecture is extensible, with a published MCP protocol for building any channel integration. And Anthropic’s existing Slack integration demonstrates they know how to ship chat-platform AI integrations at enterprise scale.

The research preview label is honest — expect API changes in the --channels flag syntax, expect the plugin ecosystem to be sparse at launch relative to where it will be in six months, and expect enterprise governance features to mature over the next two quarters. But the core capability works today on Claude Code v2.1.80 and later for teams on Pro, Max, Team, or Enterprise plans. For marketing teams who have been waiting for an AI agent that meets them where they already work, rather than one more tool they have to context-switch to access, this is the architecture worth building on now.