Sign in Subscribe
news

Claude Tag

Sebastien Dubois

Sebastien Dubois

— 8 min read

Canonical version: Claude Tag.

Claude Tag is Anthropic's way for teams to work with Claude as a teammate inside Slack. Announced June 23, 2026 (beta for Claude Team and Enterprise), it lets Claude join Slack as a team member: admins grant it access to selected channels, tools, data, and codebases, and anyone in a channel can tag @Claude to delegate tasks while they focus on other work. It replaces the previous Claude in Slack app.

Anthropic frames it as the next evolution of Claude Code: more proactive, and built for a full team rather than a single user. Internally, 65% of Anthropic's product team code is written by their internal version of Claude Tag — including most of Claude Tag itself — and the pattern has spread beyond engineering to chasing product metrics, working support tickets, and root-causing bugs.

What makes it different

  • Multiplayer: one Claude per channel interacts with everyone. Anyone can see what it's working on and pick up the conversation where the last person left off — closer to a teammate than a private chat.
  • Learns over time: it builds context from the channels it follows, so nobody has to re-explain things. With permission it can also learn from other channels and data sources (it doesn't report from private channels).
  • Takes initiative: with "ambient" behavior enabled, it proactively flags relevant information and follows up on threads or tasks that went quiet without resolution.
  • Works asynchronously: set it a task and move on. It can schedule tasks for itself and pursue a project autonomously over hours or days — enabling delegation to many Claudes in parallel.
  • DMs: message it directly and it responds privately, using your personal tools and connectors. Since September 2026 your personal connectors also work for requests you make in a channel (see #Personal connectors in channels).

Identity, permissions, and memory model

  • In channels, Claude acts under its own service accounts, which Anthropic calls "agent identity". It posts as the Claude app, opens pull requests as the Claude GitHub App, and uses an admin-provisioned service account in every other tool. Admins set a workspace baseline (connections, repositories, skills and plugins, standing instructions) that every channel inherits, and can override it per channel. Each private channel gets its own identity; public channels share the workspace-level one. Memories stay scoped: a sales Claude won't pass memories to an engineering Claude, nor leak data across scopes.
  • Two memory tiers: a workspace-wide shared tier that carries across channels, and a per-channel tier that stays local. Channel-tier memory is visible to everyone in the channel, which acts as a safeguard — anyone can see and revert a bad saved rule.
  • Admins set token spend limits (org-wide and per channel) and get an audit log of everything @Claude did, along with who requested each task.
  • Code runs in an isolated, ephemeral cloud container cloned per session — installed dependencies and uncommitted files don't persist; anything worth keeping must be committed and pushed.
  • What Claude can reach in a channel is the same for everyone who asks there. A one-to-one DM is different: it runs on your own claude.ai account, with your connectors (MCP servers included), results attributed to you (pull requests excepted) and usage billed to your seat instead of the organization. At launch, personal connectors only applied in those DMs. Since September 24, 2026 Claude can also use them for a request you make in a channel or a group DM, after you allow it; nobody else can use them (see #Personal connectors in channels).
  • Against in-channel manipulation, it applies a trust model: it distinguishes genuine directives from banter, won't persist harmful or service-degrading instructions, and confirms before destructive actions. It also distinguishes a conversation's initiator from incoming participants, and waits for resolution when people disagree rather than being commandeered.

Personal connectors in channels

Announced September 24, 2026. Until then, Claude in a channel could only use the connectors an admin had attached to it, and most organizations keep that list short on purpose: they want access to follow the person, not the channel. The catch is that most of what people need sits behind their own login (their calendar, their open deals, a plan only they can open).

Now Claude can use YOUR connectors (the ones on your claude.ai account) for a request YOU make in a channel: your calendar, your drive, your assigned accounts in the CRM, your staging deploys. Nobody else can use them. If a teammate asks the same thing, Claude can offer to use the connectors on THEIR account, never yours.

In the June agent identity post, Anthropic's argument was that with no personal credentials in play, a shared channel could never become a side door into someone's private documents. This feature brings personal credentials into channels, but only for the requester's own task, behind an approval prompt, with a screen on what gets posted.

How it works

  • Example from the announcement: Priya asks @Claude in #checkout-migration (a channel connected to GitHub) to check her Google Drive plan against what shipped. Claude reads the merged pull requests through the channel's GitHub connector, reads the doc through Priya's own Google Drive connector (separately from the channel's work), then posts what shipped and what's still open. Before this change, it would have stopped at the doc.
  • Approval first. The first time a task needs one of your connectors, Claude shows a prompt in the thread that only you can see: Allow (auto mode), Allow with review, or Don't allow. Tick "Use this choice for future requests" to stop being asked. You can change that later from the Home tab of the Claude app in Slack (Auto mode, Ask every time, Allow with review).
  • You control what gets posted. With Allow with review, Claude holds every result until you approve it. In auto mode it posts directly, except results that a sensitive-content check flags. That check looks for credentials and secret keys, personal identifiers, HR and compensation records, personnel changes, customer and deal information, privileged legal material, unannounced plans and strategy, protected-class information, medical information and security incidents.
  • Others can't steer your task. Requests other people make in your task's thread run with the channel's connections. Their messages reach Claude as information about the task, not as instructions, so a message asking it to change course or share what it found can't use your connectors.
  • Audit trail follows the person. Everything Claude does through your connector shows up in that tool's own log under your account, the same as DM work. The channel's own work stays under its service account. Claude's detailed work lives in a session only you can open; the channel sees the result.
  • Nothing to install, easy to stop. Connectors are managed under Customize > Connectors on claude.ai, and you can disconnect one at any time. A Stop button (visible only to you) under Claude's message ends a running task that uses your connectors.

Limits

  • Personal connectors don't run unattended. Scheduled routines and anything Claude starts on its own use the channel's connections, never yours. So tools needed for unattended work, or that the whole channel relies on, belong in shared connectors (e.g. an #on-call channel with shared access to the runbook, monitoring and deployment history, so Claude can triage CI issues after hours).
  • A channel that relies only on personal connectors fits closely supervised work, e.g. drafting an RFP response that pulls from pricing data not provisioned to the whole channel.
  • Whatever Claude posts is visible to everyone in the channel. Anthropic's own docs call the sensitive-content check "a screen, not a guarantee", and it doesn't consider whether other channel members have the same access you do. For anything sensitive, use Allow with review.

Admin controls and availability

  • Anthropic's pitch to admins: share a set of tools under an agent identity, let members rely only on their personal connectors (and the role-based access they already have), or decide tool by tool.
  • The settings live in the Personal connectors section of the Claude Tag admin settings and apply to every workspace and channel. On Enterprise, a "Require human review of every message" switch makes Claude hold every result for the requester (auto mode disappears). Owners can add their own sensitive topics, e.g. "Board meeting notes are confidential". Team plans have no such switch.
  • There is NO setting to turn personal connectors in channels off for an organization. To keep a service out of channels, admins have to manage the connector itself in the organization's connector settings.
  • Availability: on for every Team organization with nothing to set up. Enterprise follows; its admin settings page shows when the feature turns on. Claude Tag itself is still in beta.

Best practices (Thariq, Claude Code team)

  • Each Claude in each channel is different: "introduce" it to a new channel with a pinned message and instructions to remember (persona, when to respond, ...). This lands in its memory — think of it as its CLAUDE.md.
  • Create a personal channel (e.g. #yourname-claude) where you tag Claude for your own work with instructions specific to you; forward messages (like bug reports) there so your Claude handles them the way you like.
  • Ask Claude to maintain a pinned message it keeps updated with the status of everything, to avoid drowning in threads.
  • Have it react to top-level threads with status emojis (⏳ ✅ ❓ 🛑) so state is visible at a glance.
  • Get creative with dedicated channels — e.g. a "scheduling" channel where people tag Claude to find time on calendars.

Slack's official account called this thread "a masterclass in setting up stateful agent workflows".

How Anthropic uses it internally

  • Incident response: tag Claude in the incident thread; it pulls graphs, diffs the deploy, and comes back with root cause and the author tagged. The team approves in-thread; Claude opens the fix, lands it, watches the metric recover, and resolves the page.
  • Bug triage: Claude sits in the feedback channel and picks up reports automatically — finds the code path, reproduces, git-blames, writes a fix, and tags the owner. Only code review remains before it merges the PR.
  • Dependent work: hand it work blocked on something else (e.g. wiring up the frontend once the backend ships). It waits, watches, and shows up days later with the PR adjusted for whatever changed in review.
  • Background watchers: give it a threshold instead of a dashboard ("ping when CI stays red too long"). It stays quiet until the threshold is crossed, then posts with the failing test and culprit commit attached.
  • Launch and metric watching: point it at an A/B test with the metric and guardrails; it flags when a guardrail moves, accepts mid-run corrections, and pings when the result is significant — rollout PR ready.

Reception

  • The multiplayer aspect was widely seen as the real differentiator versus single-user agents.
  • Concerns raised on Hacker News: coworkers derailing an in-progress session, privacy when feeding information to a shared agent (mitigated by per-channel Claudes and per-channel access config), and machine-identity attribution — e.g. GitHub actions go through a single app identity, so audit logs can't distinguish which channel's Claude acted.
  • Skeptics called it "a brand name for tagging a Slack app" and noted Anthropic's rapid-fire product launches.
  • Compared to ChatGPT Workspace Agents: similar concept, but admin-configured per channel, more technical setup, and always named "Claude".
  • It fits the broader "loop engineering" trend — delegating recurring work to standing agent loops with shared knowledge layers, logging, and verification (open-source templates and skills for finding delegation opportunities appeared the same week).

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe