Claude Tag: What It Is and How to Use It in Slack

Claude Tag: The Persistent Async AI Coworker You @-Mention in Slack

A teammate types @Claude pull last week's signups and flag anything weird into a Slack channel, then closes the laptop. An hour later the channel has a thread: the numbers, two anomalies, and a follow-up question. Nobody opened a chat window. Nobody re-pasted context. The agent was already in the room.

That is Claude Tag, and it is not a chatbot.

Claude Tag is Anthropic's persistent, asynchronous AI teammate that lives inside a Slack channel. You delegate to it the same way you delegate to a coworker - @Claude, a request in plain language - and it works through the task with its connected tools, on its own time, and posts back in a thread. It launched in beta for Claude Team and Enterprise customers on June 23, 2026, and it replaces the old Claude in Slack app.

Here is the decode. Through AIBC's harness lens, Claude Tag is a persistent agent with its own loop, its own memory, and its own identity scoped to a channel. That single design choice - the unit is the channel, not the user - is why it behaves like a teammate instead of a tool you open and close.

What is Claude Tag?

Claude Tag is a standing AI member of your Slack workspace that accumulates context, takes initiative, and works asynchronously. Anyone in a channel can tag it, and it shares one identity per channel so the whole team sees the same work.

The reason the channel is the unit matters. A normal AI chat is a blank-slate session bound to one person. Claude Tag is bound to a channel, so its memory and its work are shared by default. When a colleague checks in, they pick up exactly where the last person left off, because they are talking to the same Claude.

ORG  (admin sets scope: tools, data, channels, spend limits)
│
├─ #support      ->  Claude A  ──  memory A      tools: Zendesk, Docs
│
├─ #data-team    ->  Claude B  ──  memory B      tools: Warehouse, BI
│                         :
│                         : cross-channel read (only with permission)
│                         v
└─ #leadership   ->  Claude C  ──  memory C      tools: Dashboards

One Claude per channel. Separate identities = isolated memories.

How Claude Tag works inside Slack

You @Claude a request, it breaks the task into stages, runs those stages with the tools it has access to, and replies in a thread. The loop is the same agentic loop you already know from Claude Code, moved into a multiplayer surface.

@Claude  "pull last week's signups and flag anything weird"
   │
   v
┌──────────────────────────────────────────────┐
│  Claude Tag   (one identity per channel)       │
│                                                │
│   1. Plan      break the request into stages   │
│   2. Run       execute with connected tools    │
│   3. Persist   add the result to channel memory│
│   4. Reply     post back in the Slack thread   │
└──────────────────────────────────────────────┘
   │
   v
Thread: the numbers + 2 anomalies + a follow-up question
(async - you already closed the laptop)

Tag it with something real - "write a pull request for the rate-limit bug", "pull sales numbers for Q2", "run this data analysis" - and it plans the steps and executes them in turn. The work is visible in the channel as it goes, not buried in a private DM.

Three things make this different from a request-response bot:

• Shared identity. Within a channel there is one Claude that interacts with everyone. Work is collective, not per-person.
• Asynchronous by default. Set it a task and move on. It can also schedule tasks for itself and pursue a project autonomously over hours or days.
• Tool-grounded. It acts through the tools an admin connected, so it pulls real data and ships real artifacts instead of describing them.

Claude Tag persistent memory: how it learns your company

Yes, Claude Tag remembers across conversations. As it follows along in a channel, it builds context about the work happening there, so you stop re-explaining the same background. That is the core split from a blank-slate chatbot.

It can also reach beyond a single channel. With permission, Claude can automatically gather facts from other channels and connected data sources to answer a question well. Admins control that scope, and you can run separate Claude identities with isolated memories for different use cases - one for the support channel, one for the data team - so context does not bleed where it should not.

For builders, this is the persisted-state idea applied to a team surface. The agent's behavior is a function of accumulated channel state, not whatever happens to be in one prompt. If you have wrestled with agent memory architecture, this is that problem solved at the org layer instead of the script layer.

What is ambient mode in Claude Tag?

Ambient mode is an opt-in setting where Claude posts on its own initiative instead of waiting to be tagged. With ambient behavior enabled, it proactively surfaces relevant information from the channels it monitors and the tools it is connected to, and follows up on threads or tasks that have gone quiet without resolution.

Turn it on for channels where a nudge adds value - an incident channel that should never let a thread die, a pipeline channel where a stalled deal needs a poke. Leave it off for tightly scoped or noisy channels where an unprompted agent would just add noise. Anthropic frames ambient behavior as something you enable, so treat it as a deliberate choice per channel, not a default to fight.

Proactivity is better understood as a dial with three settings than an on/off switch:

1. Tag only. Claude speaks when spoken to.
2. Scheduled. Claude runs its own tasks on a cadence you set, like a daily brief or a weekly report.
3. Ambient. Claude jumps into a thread when it judges it has relevant context to add.

Anthropic says it tuned this carefully for one blunt reason: the worst thing in the world is a bot that responds to everything with annoying context. When Claude gets that judgment wrong, you tell it so in the channel and it updates its memory, rather than you hunting through settings.

Claude Tag vs Claude Code: which agent for which job

Claude Code is a single-user agent in your terminal. Claude Tag is a multiplayer agent in your Slack channel. Same model family, different surface, different job.

        CLAUDE CODE                          CLAUDE TAG
   ┌────────────────────┐            ┌────────────────────────┐
   │  you  ->  terminal │            │  team  ->  #channel     │
   │  one session       │            │  one shared identity    │
   │  you watch it run  │            │  it runs async, alone   │
   │  session memory    │            │  persistent channel mem │
   └────────────────────┘            └────────────────────────┘
     heads-down building               team tasks you hand off

Use Claude Code when you are heads-down building and want a tight feedback loop you control. Use Claude Tag when a task belongs to a team, can run without you watching, or needs the shared context that already lives in a channel. They are not replacements for each other. If anything, the same engineering instincts apply to both - see our Claude Code guide for the single-player side and the harness engineering guide for the loop-and-memory thinking underneath both.

Claude Tag vs the old Slack app

Claude Tag replaces Anthropic's existing Claude in Slack app, and the difference is architectural, not cosmetic. The old app was a per-user request-response integration. Claude Tag adds a shared per-channel identity, persistent channel memory, asynchronous and self-scheduled work, and opt-in ambient posting.

On migration: administrators can opt in within 30 days. Note what the official announcement actually says - it is a 30-day window to migrate, not a hard "switches off on this date" deadline that some early coverage implied. If you run the old app today, the action item is to have an admin start the migration rather than wait.

How Anthropic uses Claude Tag internally

Anthropic ships a product when it has taken over the company's own week, not when a roadmap date arrives. Claude Code went out after roughly half the company was using it weekly. Claude Tag went out after it was opening most of the engineering team's pull requests.

The number, from Lamis Mukta of Anthropic's applied AI team, stated twice on the AI Native Dev podcast: 65% of Anthropic's product engineering PRs are opened by Claude Tag. That was measured on internal usage, before the public release.

One more figure from the same conversation, and it is the more useful one to copy: the share of work Anthropic is comfortable delegating to Claude went from about 30% to about 60% in a year, across every team. Share of PRs only means something if your work is code. How much of our work would we hand over is a question a support team or a marketing team can answer too, and it moves before the PR number does.

What that looks like outside engineering, from the same source:

• Marketing built an ad and copy generation pipeline. The story Anthropic tells about it is someone who started the day googling what a terminal is, and ended it having cut a 30-minute workflow down to 30 seconds.
• Sales gets a scheduled weekly brief where Claude reads what everyone did that week and assembles the stats and dashboard updates, so nobody builds the meeting deck by hand.
• Incident response uses Claude to triage, loop in the right people, and start diagnosing code problems before a human is awake.

Worth sitting with: a second company reports almost the same number. Sierra has published that 70% of its PRs are opened through its internal agent, across 75,000+ sessions for 600+ people. Two companies, two different agents, no shared methodology, so treat the agreement as weak corroboration rather than proof. It does put a shape on the target though. Two thirds of PRs is what the frontier of this looks like in mid-2026, not 100%.

Both sets of figures are self-reported and unverified.

Agent identity: why Claude Tag does not borrow your permissions

This is the architectural change that makes multiplayer work, and it is the detail most coverage skips.

When you run Claude Code or a coding copilot, the agent assumes your permissions. It works with your API keys and your access, and everything it does is recorded as you doing it. Fine for a single-player tool.

Claude Tag is given its own identity: its own permissions, its own keys for connected services. It is not acting on behalf of one person, it is acting on behalf of the team, and the audit trail says the agent did this as itself rather than impersonating a human. Anthropic names this as the key change it had to make to enable multiplayer behavior.

Permissioning then hangs off the channel rather than the org:

That design is what makes the cultural half safe. Anthropic works unusually publicly in Slack on purpose, because agents can then connect dots across an organization that no individual has visibility over. Working in public only scales if the permission boundaries are real, which is exactly what the per-channel scope is for.

If you are rolling this out: scope the channel before you invite the team in, not after.

How to set up Claude Tag

Setup is an admin task in Slack, and the important work is scoping access before you turn the team loose.

1. Confirm your plan. Claude Tag is beta-exclusive to Claude Team and seat-based Enterprise. If you are on Pro, you do not have it yet.
2. Pair Claude Tag in admin settings. An administrator connects it to your Slack workspace, replacing the old Claude in Slack app (you have a 30-day window to migrate).
3. Scope tools, data, and channels. Specify which tools and information each Claude identity can access, in which channels. Run separate identities with isolated memories for different teams.
4. Decide cross-channel reads. Choose whether a given Claude may gather facts from other channels and data sources, or stay sealed to its own.
5. Set spend limits. Configure token spend limits at both the organization and individual-channel level before usage scales.
6. Choose ambient per channel. Leave ambient off for scoped or noisy channels; enable it where proactive flagging and follow-ups earn their keep.
7. Run a real first task. Have someone @Claude an actual job and review the in-thread stages and result, so the team learns what good delegation looks like.

Four setups worth copying on day one, all of them in use inside Anthropic: a daily brief covering the last 24 hours across time zones, so you do not open the morning on a wall of unread threads; a live ping for anything urgent in the channels that matter to you; a team weekly report on what people learned that week, which scales knowledge sideways instead of leaving it in DMs; and asking Claude to build you a small piece of software, a personal dashboard tracking one workflow, deployed and shareable with the team.

Claude Tag pricing, credits, and spending limits

Claude Tag is included for Claude Team and seat-based Enterprise customers in beta, and usage is governed by admin-set token spend limits at both the organization and channel level. Anthropic has not published a standalone per-seat price for it; it ships inside those plans during the research preview.

Here is the honest part. A shared, asynchronous agent that can schedule its own work over hours or days is harder to budget than a tool one person opens and closes. Cost is now a channel-level property, not a personal one. The org and per-channel spend limits are the real control surface - treat them as a first-class setup step, not an afterthought, because "Claude has been quietly working on a project for two days" is a great feature and a surprising invoice at the same time. If usage discipline is your concern, our notes on keeping agent costs down carry over.

The SaaS repricing debate, without the hype

Claude Tag landed with a loud take attached: that a persistent AI teammate threatens the per-seat SaaS model. Andrej Karpathy called it the third major redesign of LLM UI/UX - first the LLM was a website you visit, then an app you download, now a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside humans. As framing, that is sharp and worth sitting with.

As a prediction, slow down. The "SaaSpocalypse" version of the story overshoots in three ways:

• Beta reality. It is a research preview, Team and Enterprise only, Slack only, with wider platforms still weeks out. That is not yet a category-killer.
• Pricing is not solved, it is moved. Going from per-seat to per-outcome sounds clean until you try to forecast an agent that schedules its own multi-day work. The cost-tracking problem does not disappear; it changes shape.
• The trust layer is missing. The loudest builder critique is that everyone is excited and nobody has built the layer that proves what a shared async agent did, why, and whether it was right. Until that exists, regulated and high-stakes teams will keep a human in the loop.

The signal is real: agents that hold context and act over time are a genuinely new interaction model, and Karpathy is not wrong that it reorders how software gets used. The overstatement is treating a Slack beta as the end of per-seat software. Watch the trust and cost-tracking gaps - that is where this either matures or stalls.

Is Claude Tag worth it for your team?

It is worth turning on now if your team already lives in Slack and you have real async, multi-person work to hand off. Skip it for now if you cannot scope it tightly or cannot stomach a beta on sensitive workflows.

Good fit if you:

• Already run work in shared Slack channels
• Have repeatable team tasks worth delegating (reporting, triage, analysis, PRs)
• Can set and respect org and per-channel spend limits
• Are on Team or seat-based Enterprise

Skip it for now if you:

• Are on Pro (you do not have access yet)
• Work in regulated or high-stakes flows where an unverified async agent is a liability
• Cannot scope channel, tool, and data access cleanly
• Need predictable, fixed costs more than you need autonomy

Frequently Asked Questions

What is Claude Tag? Claude Tag is Anthropic's persistent, asynchronous AI teammate that lives in a Slack channel. Anyone in the channel can @Claude to delegate a task, and it works through the task in stages with its connected tools, then posts the result in a thread. It launched in beta on June 23, 2026.

How does Claude Tag work in Slack? You @Claude a request in plain language. It breaks the task into stages, runs them with the tools it has access to, can work asynchronously and schedule future work, and replies in a thread. One shared Claude lives per channel, so the whole team sees and continues the work.

Does Claude Tag remember context across conversations? Yes. It accumulates channel-scoped memory from the ongoing discussion and connected data, so it learns your company over time. With admin permission it can also gather facts from other channels; otherwise its memory stays scoped to its own channel. It is not a blank-slate session.

What is ambient mode in Claude Tag? Ambient mode is an opt-in setting where Claude posts on its own initiative - surfacing relevant information and following up on threads or tasks that went quiet - without waiting to be tagged. Enable it per channel where proactive nudges help.

What is the difference between Claude Tag and Claude Code? Claude Code is a single-user agent in your terminal. Claude Tag is a multiplayer, persistent teammate in Slack with shared identity, channel memory, and async work. Different surfaces for different jobs, not replacements.

Does Anthropic use Claude Tag internally? Yes, heavily, and it was the release signal. Anthropic's applied AI team reports that 65% of its product engineering PRs are opened by Claude Tag, measured before the public launch, and that the share of work the company is comfortable delegating to Claude went from roughly 30% to 60% over a year. Those are Anthropic's own figures, self-reported and unverified.

Does Claude Tag use my permissions and API keys? No. Unlike Claude Code or a coding copilot, which assume your permissions and act as you, Claude Tag is given its own identity with its own permissions and keys. It acts on behalf of the team rather than one person, which also means the audit trail records the agent, not an impersonation of a human.

How much does Claude Tag cost? It is included for Claude Team and seat-based Enterprise customers in beta. Usage is governed by admin-set token spend limits at both the organization and channel level. Anthropic has not published a standalone per-seat price during the research preview.

Is Claude Tag available on Claude Team and Enterprise plans? Yes. It launched in beta for Claude Team and seat-based Enterprise customers, paired through admin settings. Anthropic plans to bring it to other platforms in the coming weeks.

How is Claude Tag different from Anthropic's old Slack app? The old per-user Slack integration is being replaced. Claude Tag adds a single shared identity per channel, persistent channel memory, asynchronous and self-scheduled work, and opt-in ambient posting. Admins have a 30-day window to migrate.

Will Claude Tag disrupt per-seat SaaS pricing? It is a real signal toward outcome-based pricing, and Karpathy called it a third UI paradigm, but the "SaaSpocalypse" framing is overstated. It is a Slack-only beta, the cost-tracking problem moves rather than vanishes, and the trust layer that would prove what a shared async agent did is still unbuilt.

Claude Tag is a beta and a research preview, so expect the access tiers, the platforms it runs on, and the spend controls to shift in the coming weeks - verify the current state against Anthropic's announcement before you wire it into anything load-bearing.