Claude Tag, explained: the AI teammate you @-tag into Slack

What did Anthropic actually launch today?

Claude Tag, launched today (June 23, 2026) in beta, is Anthropic’s new way for teams to work with Claude. Claude joins Slack as a teammate: you @-tag it into a channel, hand off a task, and it works asynchronously on Claude Opus 4.8. It is available to Claude Enterprise and Claude Team customers, and it starts on Slack.

I’ve spent a lot of the last year wiring agents into real production systems, so my first reaction wasn’t the marketing line — it was “okay, what’s actually new here, and what should a careful team watch out for?” Let me walk through both.

You can read the full details in Anthropic’s announcement.

How does it actually work?

You tag @Claude into a channel with a request. It breaks that request into stages and executes them sequentially using the tools it has access to. That last part matters: it only acts within the tools, data, and channels an admin has permitted.

The piece that makes it feel less like a chatbot and more like a teammate is the async autonomy. You can set Claude a task and walk away. Anthropic says it can “schedule tasks for itself, pursuing a project autonomously over hours or days,” and it can plan tasks for the future. That’s the difference between asking a bot a question and delegating a job. If you want the longer version of why that distinction matters, I wrote about what makes something an AI agent rather than a chatbot, and separately about agents versus plain automation tools.

The way an agent actually reaches your tools and data — calendars, docs, databases — is usually some flavor of permissioned connection. If you’re curious how that plumbing works in general, here’s my guide on connecting an AI agent to your data with MCP.

What makes it different from a normal Slack bot?

Three traits separate Claude Tag from the old “type a prompt, get a reply” pattern.

Multiplayer is the one I find genuinely clever. Most AI assistants are per-user. Here, one Claude lives in the channel, and the whole team shares its progress. That maps to how teams actually work — someone hands off, someone else picks up, the context survives the handoff.

The ambient mode is the one I’d be most deliberate about. It’s off unless you enable it, and that’s the right default. An always-on assistant that proactively flags things across channels is useful right up until it misreads a thread and proactively says the wrong thing in front of the wrong people.

Who controls what it can see and do?

This is the part every admin should read twice. The governance levers Anthropic provides:

• Administrators control which tools, data, and channels @Claude can access — and they do it per channel, not just org-wide.
• Admins can create separate Claude identities for different uses.
• Token spend limits apply org-wide and per-channel, with full activity logging.

My honest take, clearly as opinion: per-channel scoping and full logging are exactly the right primitives. If you’re going to put an autonomous agent inside your company’s conversations, you want to be able to box it into one channel, give it its own identity, cap its spend, and read back everything it did. Anthropic got those primitives right.

The caveat — also opinion — is that admin access controls are the only governance lever Anthropic highlights. That’s a real boundary, and I’ll come back to why it matters.

What’s the migration from Claude in Slack?

Claude Tag replaces the existing “Claude in Slack” app. This isn’t an add-on; it’s a swap. Anthropic gives admins a 30-day migration window to move over. Reporting says the old “Claude in Slack” app goes away on or around August 3, 2026 — treat the 30-day window as the firm fact and the exact date as “as reported.”

My read on the contrast, stated as opinion: the new app is more capable and more autonomous than what it replaces. The old app was something you prompted; Claude Tag is something you delegate to, share across a channel, and let run async. That’s a meaningful jump, and it changes how you should think about access.

Practical advice for admins: plan the cutover now, and re-set per-channel permissions deliberately. Don’t just port your old habits across. Because Claude Tag is more autonomous, “whatever access we gave the old app” is the wrong default. Start from least privilege and add access where there’s a clear need.

Is Claude Tag actually any good?

Here’s the dogfooding signal that got my attention, plus the concerns a fair post has to name. The number first: at Anthropic, 65% of their product team’s code is created by their internal version of Claude Tag.

That’s a meaningful signal. Companies that ship the most aggressive AI claims often don’t run those systems on their own critical path. Anthropic putting it on their product team’s code is the kind of dogfooding I respect.

Now the honest concerns. Claude Tag persistently learns from your company’s Slack conversations, and it can gather facts from across the org if it’s permitted. Reporting — TechCrunch’s framing, “Claude Tag is learning your company, one Slack message at a time,” captures it well — raises questions Anthropic’s announcement doesn’t fully answer: what does it retain long-term, how is that retained data stored, and what are the deletion policies? Those aren’t gotcha questions. They’re the standard governance questions you’d ask before any system starts accumulating institutional memory from your private conversations.

The always-on ambient mode adds its own risks: misinterpretation, false confidence, and unintended disclosures across channels. An agent that proactively surfaces “relevant” information can surface it to the wrong room.

My opinion, stated as opinion: the upside is real for teams drowning in async work — an agent that plans, works for hours, and is shared by the whole channel is genuinely useful. But “beta” plus “learns from your Slack” is exactly the combination that argues for scoping permissions tightly and auditing the logs from day one. The capability is impressive; the discipline around it has to come from you.

Where does this sit in the bigger AI-at-work race?

Claude Tag is part of an industry-wide push toward what’s being called “organizational context” — giving AI durable access to how a specific company actually operates. Reporting names the usual contenders in that race: Microsoft Graph and Copilot, Snowflake, Databricks, and Glean. Media is framing Claude Tag as a “virtual employee” or always-on AI teammate living in Slack.

And Anthropic isn’t stopping at Slack. They plan to expand to “many other places teams work.” So if Claude-as-a-teammate proves out, expect it to show up wherever your team already collaborates. If you want more on this whole category, I keep a running set of AI engineering guides on agents here.

FAQ

Is it generally available? No. It’s in beta, for Claude Enterprise and Claude Team customers, starting on Slack.

What model does it use? Claude Opus 4.8.

Do I lose “Claude in Slack”? Yes. Claude Tag replaces it. Admins get a 30-day migration window, and the old app reportedly sunsets around August 3, 2026.

Will it read all my channels automatically? It learns from the channels it follows, and it can learn from other channels and data sources only if granted permission. Admins scope access per channel.

Can I control cost? Yes. Token spend limits apply org-wide and per-channel, with full activity logging.

The bottom line

Claude Tag is a genuinely useful step: an AI teammate that plans a task into stages, works async over hours or days, and is shared per channel so the whole team can hand off and pick up. The multiplayer model fits how teams actually operate, and the dogfooding number is a real signal.

It’s also still beta, Enterprise and Team only, and it learns from your Slack. That combination means you scope permissions tightly and watch the logs. My verdict: worth piloting today — just not worth pointing at your most sensitive channels on day one.