For most people, "using Claude" has meant going to Claude: open a tab, type a prompt, read the answer, paste it back into wherever the work actually happens. On June 23, Anthropic announced something that inverts that. Claude Tag drops Claude into a Slack channel as a tagged team member you can hand work to, and it replies right where your team is already talking.
What Claude Tag actually does
An admin adds Claude to a Slack workspace and decides, per channel, which tools, data, and codebases it can touch. From there, anyone types @Claude and delegates a task, and Claude works on it in the thread. Anthropic describes four things that make it feel different from the chatbot you are used to.
It is shared. One Claude sits in the channel and everyone sees what it is doing, so it works more like a group member than a private assistant.
It remembers. Claude holds the channel's context over time and can carry knowledge between channels you allow, while staying out of private channels.
It can take initiative. If you switch it on, Claude will flag things and chase a stalled thread without being asked each time.
It works in the background. Claude can schedule its own steps and keep going on a project for hours while you do something else. Anthropic engineer Boris Cherny says each tag spins up its own sandbox that can clone a repo, write and test code, then throw the workspace away.
Claude Tag replaces Anthropic's older "Claude in Slack" app, which is being retired on a 30-day migration.
The proof point, and the catch
The number Anthropic is leading with: its own internal version already writes about 65% of its product team's code. (The company's Cat Wu framed the same figure as Claude merging 65% of that team's pull requests.) That is a striking claim, and it is also the tell. The flagship use here is software work, and the public release is a beta limited to Team and Enterprise accounts, not something an individual can switch on today.
Anthropic built in the controls you would want before letting an AI act in a shared workspace. Admins set per-channel identities, so a sales channel's Claude cannot see engineering data, cap how much it can spend, and get a full audit log of every action and who asked for it. It does not read private channels.
The reception was loud. Anthropic's Alex Albert said it "feels less like using a tool and more like managing a team," and Andrej Karpathy called it the third major redesign of how people interact with AI. Even Marc Benioff, whose Salesforce owns Slack, publicly welcomed it. The sharpest skeptical note came from engineer Gergely Orosz, who flagged lock-in: if he were a CTO, he said, he would want a Slack setup where he could swap the underlying model anytime.
Andrej Karpathy called this the third big redesign of how we use AI: first it was a website you visited, then an app you downloaded, and now it is something that lives where your team works and that you talk to like a coworker.
This is the first time Claude genuinely acts like a teammate, and it is a little wild to sit with.
The practical version is simpler than the paradigm talk: a lot of people are about to save real time by dropping Claude into a channel and seeing what it can take off their plate.