Anthropic launches Claude Tag in Slack beta with channel memory

• Anthropic shipped Claude Tag as a Slack-native beta for Claude Team and Enterprise, where claudeai's launch post says Claude joins as a team member with access to selected channels and tools, and claudeai's availability post says wider expansion is planned.
• The product pitch is shared context, not another private chat: claudeai's channel-memory post says one Claude serves a whole channel and builds context over time, while Boris Cherny's launch thread calls it proactive, multiplayer, with its own identity and memory.
• Under the hood, Boris Cherny's sandbox description says each tagged thread spins up its own isolated instance that can clone repos, run tests, and gets thrown away afterward, with separate memory and permissions per channel.
• Anthropic is foregrounding async delegation as the workflow shift, with ClaudeDevs' internal-usage thread and claudeai's internal metric post both saying its internal version now writes 65% of the product team's code.
• Early hands-on posts from trq212's setup thread, TobinSouth's usage notes, and daniel_mac8's company test all converge on the same pattern: private channels, pinned instructions, status reactions, and broad tool access matter more than the Slack mention itself.

You can open Anthropic's launch post, check Boris Cherny's explainer, and jump straight to Cat Wu's permissions guide. The oddest reveal is Cherny's note about one sandbox per thread, which makes Claude Tag feel closer to disposable agent sessions than a classic Slack bot. The most concrete workflow examples come from ClaudeDevs' incident-response demo, its bug-triage post, and its background-watchers post.

Claude Tag starts from a simple surface change. Claude sits in Slack as a participant in shared channels, not as a sidecar chat.

Anthropic's own framing breaks the form factor into four properties, shown in

:

• multiplayer, one Claude per channel
• memory that accumulates from channel history and connected sources
• ambient initiative when enabled
• asynchronous work over hours or days

That multiplayer piece is the big product difference. According to claudeai's channel-memory post, anyone in the channel can pick up where the last person left off, which is the opposite of restarting context in a DM every time.

The most useful implementation detail came from Boris Cherny, Claude Code lead at Anthropic. In Cherny's sandbox description, he says each tagged thread gets its own sandbox, capable of cloning repos, writing code, testing, and compiling in isolation.

That same post adds three boundaries that matter:

• one instance per thread
• its own memory per channel
• its own permissions per channel

Because the sandbox is discarded after the job finishes, Claude Tag's persistent state seems to live in channel memory and permission scopes, not in a long-lived compute environment. That helps explain why Cherny's product description emphasizes behavior that can differ by channel.

Anthropic is shipping a proactive mode, not just a mention-driven bot. In claudeai's ambient-mode post, the company says Claude can follow up on quiet threads and flag relevant information from connected channels and tools.

The internal examples turn that into concrete watcher patterns:

• ClaudeDevs' background-watchers post describes CI threshold monitoring that stays quiet until a failure condition is crossed.
• ClaudeDevs' A/B-test watcher post describes guardrail monitoring during experiments, then a rollout PR when significance is reached.
• Cherny's monitoring example says he has Claude proactively answer questions, draft PRs, and mark resolved threads with emoji reactions.

That is a more opinionated agent loop than most Slack assistants ship with on day one. The product assumes long-running tasks and delayed callbacks are normal.

Anthropic's own dogfooding numbers are blunt. Both claudeai's internal metric post and Cherny's follow-up say the internal version produces 65% of the product team's new code.

The accompanying thread gives a better map of what that number is supposed to mean in practice:

1. ClaudeDevs' incident-response demo has Claude pull graphs, diff deploys, identify a root cause, open the fix, watch recovery, and resolve the page.
2. ClaudeDevs' bug-triage post has it reproduce reports, trace the code path, git-blame the change, write a fix, and tag the owner.
3. ClaudeDevs' dependent-work example has it wait on blocked work for days, then show up with an adjusted PR.
4. ClaudeDevs' postmortem example has it rebuild a timeline from the thread, write the doc, and file action items.
5. ClaudeDevs' A/B-test watcher post extends the same loop to metrics and rollout decisions.

The throughline is that Claude Tag is being pitched less as chat and more as delegated background work with a thread as the control plane.

The memory model is where outside users started developing tactics fastest. In trq212's setup thread, Thariq Shihipar calls each channel's Claude different, then in his pinned-instructions post says he introduces a new channel with a pinned message so those instructions become part of memory, like a channel-level Claude.md.

The rest of that thread turns memory into operating patterns:

• Shihipar's personal-channel setup uses a private #name-claude channel for individual preferences and forwarded work.
• his status-board tip asks Claude to maintain a pinned status message summarizing all active threads.
• his emoji-status tip adds ⏲️ ✅ ❓ 🛑 reactions for glanceable thread state.
• his scheduling-channel idea treats channel specialization as a feature, not a workaround.

Other early testers echoed the same shape. daniel_mac8's company test says a private personal channel is the move, while TobinSouth's usage notes says Claude Tag is most useful when different channels are treated differently.

Availability is narrow for now. claudeai's availability post, ClaudeDevs' beta note, and Cherny's rollout post all place Claude Tag in beta on Slack for Team and Enterprise customers.

The setup burden also looks real. TobinSouth's admin note says the system becomes much more capable once admins connect GitHub repos, Slack, internal knowledge bases, and MCPs, while Cherny's security overview says Anthropic is layering safety across model training, classifiers, access controls, and workspace boundaries.

A few reply-level details add texture that the launch posts skip:

• Cherny's "Soon!" reply says more surfaces are coming.
• Cherny's confirmation reply confirms a user question, indicating at least one requested capability is already supported.
• Cherny's commit-attribution reply says commit attribution settings can be changed just by asking Claude to update them.

That last point is small, but it is new product surface area: some configuration appears to be conversational, not hidden behind an admin panel.