Claude Tag: Claude now works inside your Slack as a teammate
Anthropic launched Claude Tag on June 23, 2026 — a version of Claude that lives inside Slack as a shared teammate you hand work to by typing @Claude. It's a research-preview beta for Claude Enterprise and Team plans, runs on Opus 4.8, and replaces the old Claude-in-Slack app. Wider availability is promised later.
Here's what it actually is, where it earns its keep, and who it isn't for yet.
What it is
Claude Tag is Claude Code grown into a team setting: less "a tool one person prompts," more "a coworker the whole channel hands work to." Tag it with a request and it plans the work in stages, runs each step using the tools you've granted it, and reports back in the thread with the output — a pull request, an analysis, an incident fix.
Three things separate it from a chatbot sitting in a channel:
- It's shared. A single Claude lives in a channel and accumulates context as it follows along, so another teammate can step in and continue without re-briefing it.
- It's async. Assign something and leave. It can grind through a task over hours, or sit on work it's blocked on for days, and resurface with the result.
- It can take initiative. Switch on "ambient" mode and it acts without being summoned — nudging stalled threads and surfacing things worth knowing from the channels and tools it can see.
Where it actually helps
The strong fit is engineering and ops work that's bursty, asynchronous, and heavy on context — the stuff that usually dies in a Slack thread because nobody had time to chase it:
- Incident response. When a page fires, tag it into the thread. It gathers the relevant graphs, compares the latest deploy, and returns with a probable cause and the engineer who shipped it. Once your team signs off in-thread, it opens the fix, ships it, and waits to confirm the metric recovers before closing the page.
- Bug triage. Parked in a feedback channel, it picks up reports on its own — tracing the code path, reproducing the issue, checking who last touched it, and drafting a fix with the owner looped in. A human reviews before anything merges.
- Blocked work. Hand it a task waiting on something else — the frontend that can't land until the backend ships. It holds, watches, and comes back later with a PR adjusted for whatever shifted in the meantime.
- Postmortems. After an incident closes, ask it to write the writeup. It re-reads the thread, reconstructs the timeline, drops the doc where your team keeps them, and turns the follow-ups into tracked issues.
- Background watchers. Instead of a dashboard, give it a condition — alert if the build stays broken past some window. It stays quiet until that line is crossed, then posts the failing test and the commit that likely caused it, ready to attempt a fix on your say-so.
- Experiment watching. Point it at a running experiment — the metric you care about, the limits you can't cross. It speaks up if one of those limits drifts, lets your team correct mid-run, and flags when the result is conclusive with the rollout PR prepared.
The honest catches
- Most people can't use it yet. Enterprise and Team only — not Pro or Free. If you're an individual user, treat this as a signal of where things are heading, not something you can switch on today.
- It's a beta. Research preview means rough edges; don't wire it into anything you can't supervise.
- It acts in your workspace with real tool access. That's the value and the risk — it can open PRs, run analyses, and reach connected services. A human still signs off on the consequential steps (the merge, the fix), so treat it as delegation with review, not autonomy. And ignore the "virtual employee replacing headcount" framing some coverage reached for: it removes the chase-work, not the judgment.
- "Ambient" can get noisy. Proactive flagging helps until it doesn't. It's a setting you control, so tune it.
- Migration has a deadline. The legacy Claude-in-Slack app retires August 3, 2026, and admins have a 30-day window to opt in. Anthropic is issuing launch credits to eligible orgs.
The reassuring part — and the part worth actually setting up — is the governance. Admins scope each channel's Claude to specific tools, data, and channels, with separate identities, so the instance in a sales channel can't see what the one in an engineering channel can. Admins can also cap token spend per organization and per channel, and pull an audit log of every task Claude ran and who asked for it. That's the kind of control that makes an in-Slack agent defensible — but it only protects you if someone configures it deliberately rather than clicking through.
One number for context: Anthropic says roughly 65% of its own product team's code is now written by its in-house version of this tool. That's a real signal it's more than a demo — but it's also Anthropic dogfooding their own setup, so read it as "this is serious," not "you'll get the same on day one."
Bottom line
Claude Tag is a genuine shift: Claude moving from a tool you prompt to an async teammate you delegate to, with shared context across a channel and admin controls built for enterprises. If you run an engineering or ops team on Slack with Enterprise or Team, it's worth piloting now — scoped carefully. If you're an individual user, it's not for you yet, but it's a clear preview of how this kind of work gets done next.
Sources: Anthropic — Claude Tag · TechCrunch · VentureBeat · Cyber Security News (security/Opus 4.8 details). Verified 2026-06-23; plans, dates, and details may change.