Claude Tag

Claude Tag is a team-oriented agent product from Anthropic that places Claude inside Slack as a shared AI agent. A channel member can mention @Claude with a task, after which Claude works in a hosted sandbox, uses the tools assigned to that channel, and returns its result in the same Slack thread. Unlike a private chatbot conversation, a channel session is visible to the team and can be steered by anyone in the channel.^[1][2]

Anthropic announced Claude Tag on June 23, 2026 and released it in beta for Claude Team and Enterprise customers. Slack was the only supported team platform at launch, although Anthropic said it intended to bring the product to other work environments later. The launch version used Claude Opus 4.8.^[1] Claude Tag is not a separate language model. It is a collaborative agent interface and execution system built around Claude.

Scroll sideways for more →

Release and product positioning

Claude Tag extended an earlier Claude in Slack experience from conversational assistance into longer-running agent work. Anthropic said the older Claude in Slack experience would switch to Claude Tag on August 3, 2026. The new system retained channel mentions, direct messages, and the Slack assistant panel, but added organization-owned identity, shared memory, scheduled work, proactive follow-ups, and access to organization tools.^[2]

Anthropic described the product as the start of an evolution of Claude Code. The company said its internal version created 65 percent of the code produced by its product team, a figure reported as Anthropic's own measurement rather than an independent audit. It also described internal use for product metrics, support tickets, data analysis, and debugging.^[1][11]

The product name refers to the act of tagging @Claude in a work conversation. Anthropic consistently capitalizes the name as "Claude Tag"; @Claude is the Slack identity that participants mention.

How a channel session works

Mentioning @Claude with a request starts a working session for that Slack thread. Anthropic creates an isolated cloud sandbox for the session. Claude reads the thread and channel context, builds a checklist for longer tasks, calls tools that the channel is permitted to use, and posts the result back into the thread. Results can include a Slack reply, file, chart, document, hosted page, or draft pull request.^[3]

The sandbox is temporary, while the Slack thread is durable. When activity stops, Anthropic releases the sandbox. A later reply rebuilds a new sandbox and resumes from the thread context. Files that existed only inside the earlier sandbox do not survive, so durable work must be posted to Slack, pushed to a repository, or saved in another connected system.^[3][5]

Each thread has a separate session and sandbox. Two threads in the same channel do not share their active filesystem state, although they can use the same channel tools and shared memory. The execution engine is the same managed system used by Claude Code on the web. This allows Claude Tag to run code and produce working artifacts without giving it access to a user's local computer.^[3]

Shared team behavior

Claude Tag treats a channel thread as a shared task rather than a private exchange. Anyone who can participate in the channel can see the work, reply with more context, redirect the approach, or continue after the original requester leaves. Once a session is active, a reply in the thread is sufficient; another @Claude mention is not required.^[3]

Edits and deletions behave differently from new replies. Claude receives a note when an earlier message is edited but does not treat the edit as a new instruction. It receives no notice when a Slack message is deleted and may retain the version it already read. A participant who needs to correct the task must add a new reply.^[3]

Anthropic also introduced proactive and asynchronous behavior. A channel can use routines triggered by a schedule, a channel watch, or a repository event. If ambient behavior is enabled, Claude may surface information from permitted channels and tools, report that work finished, or follow up when a thread has stalled. Scheduled work uses the same channel identity and access as a task started by a person.^[1][3]

Channels, direct messages, and the assistant panel

Claude Tag is available through three Slack surfaces, but they do not all use the same identity or data access.^[2]

Scroll sideways for more →

Direct messages are separate from the organization's channel agent. Work in a DM is attributed to the individual, and the organization-provisioned service accounts do not apply. Group direct messages are not supported. Organization owners can disable direct messages for the deployment.^[6]

This separation also distinguishes Claude Tag from Claude Cowork. Anthropic's documentation positions Tag for shared work that a team should see and steer, Cowork for personal research and drafting in a Claude chat, and Claude Code for hands-on programming in a terminal or integrated development environment. Claude Tag uses the same type of managed agent execution as Claude Code on the web, but channel sessions use team-assigned service accounts rather than a developer's local credentials and filesystem.^[3]

Identity, permissions, and administration

In a channel, Claude acts through accounts created for the agent rather than impersonating the person who made the request. Slack posts come from the Claude app. On GitHub, branches and pull requests created through the official integration are attributed to the Claude GitHub App. Other connected services use an account created for Claude, such as a service account with its own email address.^[4][6]

Only a Claude organization Primary Owner or Owner can configure Claude Tag access and channels; the Claude Admin role can view some settings but cannot complete setup. Pairing the app also requires a Slack workspace administrator. Setup links the Slack workspace, provisions Claude's identity, grants selected repositories and tools, sets an organization spending limit, and tests the agent in a channel.^[2][4]

Access can be assigned at organization, workspace, and private-channel levels. Workspace permissions apply to public channels in that Slack workspace, while a private channel can receive additional credentials or repositories. Each narrower scope inherits access from the levels above it. This means that a private-channel grant should add only the access needed by that group, while the organization-wide baseline remains limited.^[2]

Member controls differ by plan. A Team organization can leave the agent open to anyone in the connected Slack workspace or restrict it to Slack users who also belong to the Claude organization. Enterprise adds role-based access control through custom roles. These restrictions cover both channel requests and direct messages.^[2][8]

Credential and network security

Channel work runs on Anthropic's infrastructure, not inside the customer's local network. The sandbox does not hold connected-service credentials. An Agent Proxy retrieves a credential from a separate store and injects it only when an outbound request matches an allowed rule. Network egress is denied by default except for hosts allowed by a connection, an access bundle, or the session environment.^[5]

An owner can narrow a connection to a host, a path prefix, or read-only request methods. The external service sees Claude's own account, which makes actions attributable in that service's logs. Anthropic also provides an organization audit view for one-time and scheduled tasks and for network calls made through the agent identity.^[2][5]

Claude has no access to external business systems until an owner adds connections. Within Slack, however, it can read and post in channels where it has been added and can search public workspace channels by keyword under Slack's normal search permissions, even when the app is not a member of every public channel. Workspace search is unavailable in channels that contain guests. Administrators can restrict guest channels, external shared channels, member access, and the scopes where the agent is enabled.^[3][5][8]

Memory and data handling

Claude Tag keeps memory by place rather than by individual user. Notes learned in public channels can become workspace memory available to other public channels. A private channel can read the shared workspace memory, but facts learned in that private channel are saved only to its private store. Direct messages and other Slack workspaces remain separate.^[7]

Channel members can ask Claude what it remembers, correct an entry, or tell it to forget an entry. Claude organization admins can inspect memory, while owners can edit or delete it through organization settings. The memory is a curated set of notes, not a full transcript search. Claude can read past session transcripts when asked, but the user should identify a topic or timeframe because it cannot run full-text search across all earlier sessions.^[7]

Slack conversations with Claude remain separate from conversations in the Claude web app. Slack-originated conversations do not appear in Claude chat history, and Claude web conversations are not exposed in Slack. If the integration is disconnected or uninstalled, Anthropic says it deletes the associated Claude-side conversation data within 30 days. Copies retained in Slack continue to follow the organization's Slack retention policy.^[2]

Claude Tag retains channel memory and session transcripts so that work can continue across tasks. For that reason, it is not available to organizations using Anthropic's Zero Data Retention setting.^[4][5]

Supported workflows

The product can perform a task only when its channel has the required context and connections. Anthropic documents both Slack-only workflows and tasks that act in outside systems.^[1][3]

Scroll sideways for more →

For repository tasks, Claude clones only a repository granted to that channel into the hosted workspace. It can inspect code and commit history, then return a draft pull request under the Claude GitHub App. Repository instructions in CLAUDE.md can guide its work. Other source-control systems can be connected with their own credentials.^[9]

Thread-to-artifact tasks can work without an external connector when Slack already contains the needed information. If an organization connects an issue tracker such as Jira, Linear, or Asana, Claude can turn the same discussion into a filed ticket and post the link. Hosted pages are published on Claude's service and use channel-based access.^[10]

Availability, billing, and prerequisites

Claude Tag's beta is available on Anthropic's first-party Claude Team and Enterprise plans. It is not available to individual Free, Pro, or Max plans or to third-party Claude deployments. The organization must also have routines enabled and must not use Zero Data Retention.^[4]

Channel work is consumption-based and billed to the organization's usage balance rather than to each person who invokes it. Owners can set organization-wide and per-channel spending limits, receive threshold alerts, and review usage by channel. Work that would exceed a limit is declined. Direct-message work is billed to the individual's Claude account instead.^[2]

The deployment depends on several administrators and external accounts. A Claude Owner must configure access, a Slack administrator must approve and pair the app, and each connected service needs a limited account or official app installation for Claude. A connection added after a thread begins may require the user to name the service explicitly or start a new thread before all configuration changes are visible.^[3][4]

Limitations

At launch, collaborative Claude Tag sessions worked only in Slack. Anthropic described Slack as the first location for the product but did not announce dates or names for later platforms.^[1]

The beta excludes individual Claude plans, third-party deployments, and Zero Data Retention organizations. It also requires the organization's routines feature to be enabled. Direct messages are one-to-one, with no group-DM support.^[4][6]

Context is bounded. When Claude is mentioned partway through an existing Slack thread, the session receives up to 50 messages from the start of that thread after filtering replies from other bots. In a long thread, recent messages immediately before the mention may be outside that window, so critical details need to be restated. The working filesystem is also temporary, and unsaved sandbox files disappear when the session is released.^[3]

All participants in an allowed channel work with the same agent permissions, regardless of their personal access to connected systems. Anyone in the channel can also steer the active task and manage its channel memory unless plan-level restrictions block them. Organizations therefore control the product through channel membership, scoped service accounts, network rules, member restrictions, spend limits, and audit logs rather than through the requester's individual credentials.^[5][7][8]

References

1. Anthropic, "Introducing Claude Tag," June 23, 2026 ↩
2. Claude Help Center, "What is Claude Tag?," June 23, 2026 ↩
3. Claude Documentation, "How Claude Tag works," accessed July 24, 2026 ↩
4. Claude Documentation, "Set up Claude Tag," accessed July 24, 2026 ↩
5. Claude Documentation, "Security and data handling," accessed July 24, 2026 ↩
6. Claude Documentation, "How agent identity works," accessed July 24, 2026 ↩
7. Claude Documentation, "What Claude Tag remembers," accessed July 24, 2026 ↩
8. Claude Documentation, "Restrict where Claude Tag operates," accessed July 24, 2026 ↩
9. Claude Documentation, "Work with your GitHub repositories," accessed July 24, 2026 ↩
10. Claude Documentation, "Turn threads into docs and tickets," accessed July 24, 2026 ↩
11. Ross Kelly, "Meet Claude Tag, Anthropic's new AI teammate that works in Slack," IT Pro, June 24, 2026 ↩