Claude Code Channels + Auto Mode: Control Your Code from Telegram While It Runs Unsupervised [2026]

Two features in seven days. That's fast even for Anthropic.

On March 20, they shipped Channels: message Claude Code from Telegram or Discord while it runs on your machine. Four days later, auto mode: Claude decides its own file and command permissions instead of interrupting you constantly.

Separately, each one is useful but limited. Together, they turn Claude Code into something you can text from your couch and it'll keep working without babysitting. That's the pitch anyway. Here's what actually happens when you try it.

Channels is an MCP plugin that pushes messages from Telegram or Discord into a running Claude Code session. It's not a separate app or a cloud service. Your local Claude Code instance, the one with access to your actual codebase, receives messages from your phone and responds in the same chat.

According to Anthropic's docs, the flow is: you send a message on Telegram, the plugin polls for it, Claude Code processes it against your local project files, and the response appears in your Telegram chat. All code execution happens on your machine.

MacStories ran a hands-on review and described it as "very capable, but a little fiddly to set up." Federico Viticci went further in a Club MacStories piece, recreating his entire OpenClaw workflow using Channels. He ran Claude Code on a Mac mini and controlled it from his iPhone and iPad.

From the official docs and multiple setup guides:

1. Make sure you have Claude Code v2.1.80 or later
2. Get a bot token from Telegram's BotFather
3. Launch Claude Code with the channels flag:

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

1. DM your bot on Telegram
2. Approve the pairing code that appears in your terminal
3. Lock it down

That's the clean version. The reality, per multiple developer write-ups, involves running inside tmux or screen because Channels dies when your terminal closes.

tmux new -s claude
claude --channels plugin:telegram@claude-plugins-official

Discord setup follows the same pattern with a Discord bot token instead.

The biggest problem: messages vanish if your session is down. Channels has no message queue. Send a Telegram message while Claude Code isn't running, and that message is gone forever. No replay, no catch-up. The Anthropic docs say this explicitly, but as Vivek Karmarkar wrote in his Medium stress test after losing three messages on day one: "the tmux advice is not optional."

Then there's the permission stalling. When Claude hits an action that needs approval, it pauses and waits for you to confirm at the terminal. You can't approve from Telegram or Discord. Your remote session just stops until you walk back to your computer. For a "code from the couch" tool, this is a dealbreaker on its own.

Karmarkar also found a zombie process bug. A previous Claude Code session can keep running in the background, fighting the current session for the Telegram polling slot. Telegram only allows one consumer to poll at a time. Two processes competing means messages get randomly split between them. Half your messages just disappear.

And there's no delivery confirmation. You send a message and hope Claude got it. No read receipts. No connection indicator.

Here's the context. Claude Code's default behavior asks permission for every file write and every bash command. You tell it to refactor a module, and it interrupts you dozens of times: "Can I edit this file? Can I run this test? Can I save this change?" UC Irvine research (cited in Anthropic's blog) found that knowledge workers need over 20 minutes to refocus after an interruption. A single Claude Code session can generate dozens.

Developers got frustrated and started using --dangerously-skip-permissions, which removes all safety checks. Every action proceeds, no questions asked. Convenient. Also terrifying if Claude decides to do something destructive.

Auto mode launched March 24 (TechCrunch, 9to5Mac, Anthropic engineering blog). It's the middle ground. A safety classifier reviews each action before it runs. Safe actions proceed automatically. Risky ones get blocked and Claude tries a different approach.

Simon Willison covered it the same day on his blog, describing it as "a permissions mode where Claude makes permission decisions on your behalf, with safeguards monitoring actions before they run."

Anthropic published the classifier's performance numbers:

The 0.4% false positive rate means about 1 in 250 safe actions gets blocked. When that happens, Claude doesn't stop and wait for you. It retries with a different approach. Your session keeps moving. If the classifier blocks the same action 3 times in a row or 20 times total, auto mode gives up and switches back to manual prompts.

The 5.7% false negative rate is the scarier number. That means about 1 in 18 synthetic attack attempts got through the classifier. Anthropic tested this against things like credential scanning, data exfiltration, and running code from external repos. 94.3% were caught. 5.7% weren't.

TechCrunch noted that Anthropic "recommends using the new feature in isolated environments" and it currently works with Sonnet 4.6 and Opus 4.6 only. Available on Team plans now, Enterprise and API rolling out.

Channels alone has the permission-stalling problem. Auto mode alone solves permissions but you still need to be at your terminal to type prompts. Together: you text Claude from Telegram, it runs with auto mode permissions, and only interrupts you for actions the classifier flags as risky.

That's the setup Garry Tan (Y Combinator CEO) has been evangelizing. He open-sourced his Claude Code workflow on March 12, called it "gstack," and it hit 20,000 GitHub stars with 2,200 forks within days. TechCrunch covered the backlash too. Critic Mo Bitar called it "a bunch of prompts in a text file." On Product Hunt, someone pointed out: "If you weren't the CEO of YC, this wouldn't be on PH."

Fair criticism. But the underlying workflow is real. Run Claude Code in tmux with auto mode enabled, connect Channels to Telegram, and you have a coding agent you can direct from your phone that doesn't stop every 30 seconds to ask permission.

It's not magic though. The permission stall problem shrinks with auto mode but doesn't disappear. The classifier will still block some actions, and when it does, you need terminal access. And messages still vanish if your tmux session crashes.

OpenClaw wins on platforms (7 vs 2), persistence (daemon vs session-bound), message handling (queued vs lost), and price (free vs $20+/mo). Channels wins on security (Anthropic's classifier), code understanding (your actual local codebase), and setup speed.

The DEV Community comparison put it well: many developers end up using both. OpenClaw as the persistent assistant that never sleeps. Channels for focused development sessions where deep codebase understanding matters.

Both features are research previews. Anthropic says behavior may change later in 2026. They're being honest about that, which I appreciate more than companies that ship broken features and call them "GA."

Auto mode is the easier recommendation. If you use Claude Code at all, enable it. The permission interrupt problem is real and this fixes it for 99.6% of actions. The 5.7% false negative rate means you should use it in sandboxed environments, but for local development it's a clear upgrade over clicking "allow" 40 times per session.

Channels is harder to recommend without caveats. The message loss problem and permission stalling make it frustrating as a primary workflow. But as a "check in on your running task from the couch" tool paired with auto mode, it works well enough. Run it in tmux. Don't trust it with anything you can't afford to re-send.

For the full guide to Claude Cowork (the non-coding version of autonomous tasks), or a comparison of Claude Code vs Cursor for your main coding workflow, we've covered both in depth.

What are Claude Code Channels?

Channels is a plugin that connects a running Claude Code session to Telegram or Discord via MCP. You send messages from your phone, Claude processes them against your local codebase, and replies in the same chat. Launched March 20, 2026 as a research preview. Requires Claude Code v2.1.80+ and a claude.ai login.

What is Claude Code Auto Mode?

Auto mode lets Claude execute file writes and bash commands without asking permission each time. A safety classifier running on Sonnet 4.6 reviews every action before it runs. It catches 99.6% of safe actions correctly and blocks 94.3% of malicious attempts. Launched March 24, 2026.

Do Channels work when my terminal is closed?

No. There is no message queue. Any message sent while your Claude Code session is down is permanently lost. Anthropic recommends running Claude Code inside tmux or screen for always-on operation.

Can I approve permission prompts from Telegram?

No. When Claude hits a permission gate, the session stalls until you confirm at the terminal. Pairing Channels with auto mode reduces this problem since most actions proceed automatically, but the classifier will still occasionally block actions that require local confirmation.

How does auto mode compare to --dangerously-skip-permissions?

Auto mode is the safer replacement. The old flag skipped all checks. Auto mode runs a two-stage classifier on every action, blocking things like mass file deletion, credential scanning, and data exfiltration. If it blocks 3 times in a row or 20 times total, it falls back to manual prompts.

Should I use Channels or OpenClaw?

OpenClaw supports 7 platforms (including WhatsApp, iMessage, Slack), runs as a persistent daemon, has a message queue, and is free open-source. Channels supports 2 platforms, dies when your terminal closes, loses messages, and requires a $20+/mo Claude subscription. OpenClaw is more mature for persistent mobile coding. Channels is better for focused sessions where local codebase context matters.

Is auto mode safe for production?

Anthropic recommends isolated/sandboxed environments. The 5.7% false negative rate on synthetic attacks means roughly 1 in 18 malicious actions could slip through. For local development, that risk is manageable. For machines with production credentials or sensitive data, wait for the numbers to improve.

Keep Reading

AI Tools

How to Use Claude Cowork: Setup, First Task & Tips

AI Tools

Claude Code: Complete Guide to Agentic Coding

AI Tools

Claude Code vs Cursor: $20/mo Each but One Has $1,400 Overages