Claude Code Channels Add Remote Approval but Stay in Preview

Claude Code Channels can now carry tool-approval requests to a phone, removing a major obstacle to managing a live coding session away from the terminal. Telegram and Discord remain supported, iMessage has joined them, but Anthropic still labels the feature a research preview rather than a production-stable interface.

This is a meaningful change from the first release, not a new launch. The official Claude Code changelog records Channels in version 2.1.80 on March 19, 2026, permission relay in version 2.1.81 on March 20, and Console API-key authentication in version 2.1.128 on May 4.

What Channels actually connects

A Channel is an MCP server that pushes an outside event into a Claude Code session already running on the user’s machine. With a two-way messaging plugin, a Telegram or Discord message enters that existing conversation, Claude works against the files and tools available there, and its response returns through the same bot.

That architecture matters because Channels does not create a fresh cloud task for every message. The local session retains its current conversation and working context, while the messaging app acts as another route into it. A CI alert, deployment notification or monitoring webhook can use the same mechanism, so the feature is broader than phone-based chat control.

The session must remain open to receive events. Closing Claude Code, shutting down the host or losing the channel subprocess breaks the bridge; installing a plugin does not turn the computer into an independently managed cloud agent. A persistent terminal or supervised background process may keep the session reachable, but operational resilience remains the user’s responsibility.

The current feature is broader than the original Telegram and Discord release

As of August 13, 2026, Anthropic’s current Channels documentation lists Telegram, Discord and iMessage as supported plugins, alongside the local fakechat demonstration. It requires Claude Code 2.1.80 or later and Bun for the prebuilt plugins; Anthropic authentication can use a claude.ai account or a Console API key, while Bedrock, Vertex AI and Microsoft Foundry are excluded.

Availability also depends on account administration. Individual Pro and Max users can opt a channel into a session with the command-line flag, while Team and Enterprise organizations must have Channels enabled by an owner. Managed organizations can replace Anthropic’s default plugin allowlist with their own list of approved channel plugins.

This is not the same product path as Remote Control. Remote Control exposes a local session through Anthropic’s web or mobile client, whereas Channels accepts pushed events from non-Claude systems. Telegram and Discord are therefore useful when a team already works in those apps or when an automated system needs to deliver an event without opening the Claude interface.

Remote approval changes the unattended-work equation

A remote message may ask Claude to inspect a repository or continue a task, but a sensitive tool call can still trigger an approval checkpoint. Earlier setups could stall at that point until someone returned to the terminal. The permission-relay capability allows a compatible two-way channel server to forward the request and accept an allow-or-deny verdict from an authenticated sender.

The official Channels protocol reference specifies that permission relay requires Claude Code 2.1.81 or later. It covers tool-use approvals such as Bash, Write and Edit, while project-trust and MCP-server consent dialogs remain local-only; the terminal prompt also stays active, and the first valid answer from either the terminal or the remote channel wins.

That distinction makes the feature more useful without making it completely headless. A compatible channel can unblock many ordinary coding actions from a phone, but it cannot replace every local security or configuration decision. Users should therefore treat remote approval as a narrower control surface, not as evidence that any paused session can always be recovered through chat.

Pairing is the security boundary, not a convenience step

Telegram and Discord bots are externally reachable, so Anthropic’s approved plugins gate inbound messages by sender identity. The initial direct message produces a pairing code that must be approved from Claude Code; after pairing, an allowlist can silently reject other senders. Organizations add a second boundary by deciding whether Channels may run and which plugins may register.

Permission relay raises the value of that allowlist. Anyone allowed to return an approval verdict may authorize code execution or file changes within the permissions offered by the session. A bot account should therefore be protected like an administrative interface: tokens should not be exposed, obsolete sender IDs should be removed, and group access should not be enabled merely for convenience.

The flag that bypasses most permission prompts remains available for trusted environments, but it is not the necessary default now that relay exists. It also does not eliminate every prompt, and it broadens what a remotely initiated instruction may do without review. For most interactive remote workflows, preserving approvals and relaying supported ones provides a clearer control point.

A practical setup sequence

The setup has four dependencies: an authenticated and current Claude Code installation, Bun, a bot created on the selected messaging platform, and permission to use Channels under the account’s organization policy. Telegram users install the official Telegram plugin, configure the BotFather token, then restart Claude Code with the plugin named after the Channels flag. Discord follows the same overall pattern but also requires Discord-side application and bot configuration.

Install the Telegram or Discord plugin from the official Claude plugin marketplace.
Configure the platform token and reload or restart so the plugin becomes active.
Launch Claude Code with the selected plugin explicitly passed to the Channels flag.
Send the bot a direct message, approve the returned pairing code locally, and switch to an allowlist policy.
Verify ordinary replies and a low-risk permission request before depending on the bridge away from the computer.

The last check is important because several layers can appear connected while inbound events remain blocked: the bot platform, plugin process, organization setting and active Claude Code session must all agree. Testing with a harmless repository question and a deliberately reviewable action distinguishes message delivery from permission relay.

What the research-preview label still means

Channels has gained capabilities since its March debut, but Anthropic continues to warn that rollout, command syntax and the protocol contract may change. The feature’s flags may work without appearing in the standard command help, and custom channel development still uses an explicitly dangerous development flag unless the plugin is on the effective approved list.

For developers, the useful conclusion is precise: Claude Code can be reached from Telegram or Discord while its local session is running, and compatible channels can now relay many tool approvals instead of merely reporting that work has stopped. It is substantially more practical than the first iteration, yet its preview status, host dependency and remaining local-only dialogs still rule out treating it as a fully autonomous remote operations service.

Claude Code Channels Can Approve Tools From Your Phone—but Remain in Preview