Anthropic launches Claude Tag as a shared Slack agent

Anthropic launched Claude Tag in beta for Claude Team and Enterprise customers on June 23. The announcement describes a shared Slack agent that can join selected channels, receive tasks through @Claude, and use tools, data, and codebases connected by the organization. It replaced Anthropic’s earlier Claude in Slack app.

Shared context changes the authorization problem

Within a channel, Claude Tag interacts with multiple people and retains relevant channel context. Anthropic says administrators can set organization and channel spending limits and inspect an action log that records what the agent did and who requested it. The product can break work into stages, use configured tools, and return results to a Slack thread.

A shared agent must reconcile three permission sets: the requesting user, the channel, and each connected system. Granting the bot a broad credential can let a low-privilege channel member cause actions they could not perform directly. Memory also needs boundaries when membership changes or sensitive material appears in a previously broad channel.

Start with an intentionally narrow channel

Anthropic itself recommends initial testing in a private channel. Administrators should go further: connect a read-only test system, name allowed task classes, require approval for writes, and verify that audit entries link the human requester to every tool invocation. Test thread moves, deleted messages, guest users, removed members, and attempts to reference inaccessible channels.

Do not use the monthly spend limit as a security boundary; it caps cost after permissions have already been exercised. The immediate rollout artifact should be an access matrix showing which people can invoke the agent, which channel data enters context, which external actions it can take, and who reviews the resulting logs.