The GenAI Field Guide · Trending
A shared Claude agent that lives in Slack channels, works under its own service accounts, and turns a tagged request into a finished result in the thread.
Claude Tag puts one Claude per channel into Slack: anyone in the channel can tag @Claude with a task, it works in a cloud sandbox with tools an admin granted to that channel, and it posts the answer, chart, doc or draft pull request back in the thread. It replaced the earlier per-user Claude in Slack app and is in public beta for Team and Enterprise plans. Teams who want an agent their whole channel can steer, and the admins who must decide what that agent can reach, should read its access model first.
Anthropic announced Claude Tag on 23 June 2026 as a beta for Claude Enterprise and Team customers. Under the product name it is a team agent rather than a chat assistant: within a Slack channel there is one Claude that everyone talks to, it keeps notes about the channel over time, and it can take on standing work such as scheduled digests or watching for alerts. Anthropic's head of product for Claude Code and Cowork, quoted by Fortune, contrasted it with Anthropic's other tools, which she described as single-player.
The big design change from the earlier Claude in Slack app is identity. The old app answered as whichever user had linked their claude.ai account and used that person's connectors. Claude Tag in a channel acts under service accounts an organization Owner provisions, so its access is set per channel, its actions show up in each tool's audit log under Claude's own name, and code is authored by the Claude GitHub App. Direct messages are the exception: they still run on the sender's own claude.ai account.
Anthropic says 65 percent of its product team's code is created by its internal version of Claude Tag. Fortune reported the same figure differently, as Claude Tag approving and incorporating 65 percent of code changes the product team submits. Treat the number as Anthropic's internal claim either way. The documentation describes it as the same engine that powers Claude Code on the web, running in Slack, and ITPro reported Anthropic calling it the beginning of an evolution of Claude Code.
Tagging @Claude starts a session. A task that needs tools or a longer exchange gets its own Slack thread, and each thread is bound to its own session with its own ephemeral cloud sandbox. Claude posts a checklist it edits in place, works through the task, and returns a reply, a file or chart, a page kept current, a web page hosted on claude.ai, or a draft pull request. Anyone in the channel can steer a running session by replying in the thread, and a Stop button halts it. A few minutes after a turn finishes the sandbox is released; the transcript persists on Anthropic's side, and the next reply resumes the session in a fresh sandbox, so files that lived only in the sandbox are lost unless pushed or posted.
Access is the core mechanism. Admins group connections, repositories, plugins and instructions into Access bundles and attach them at one of three scopes: the whole organization, a workspace, or a single channel. The sandbox never holds the credentials. Outbound traffic is denied by default and passes through Agent Proxy, which injects a credential from a separate store only for hosts the admin allowed, sends allowlisted hosts without credentials, and blocks everything else. Connections can be narrowed to one host, a path prefix, or read-only methods, and a federated option lets Claude present a short-lived signed token to AWS, Google Cloud, a gateway or an OAuth server instead of a stored key.
Memory follows places, not people. Each channel has its own notes, and notes Claude saves from a public channel become workspace notes that every channel in the workspace reads. Anyone in a channel can ask what Claude remembers and correct or remove entries. Since late September a member can also let Claude use their own claude.ai connectors for their own requests in a channel; that work runs with their permissions and, on Enterprise, an admin can require the member to review each result before it posts.
Claude can start work unprompted when a channel's automatic replies are on, and routines run on a schedule, on a channel watch, or on a repository event, with the same channel access a typed request gets. The model is selectable: Anthropic's announcement and press coverage said it launched on Claude Opus 4.8, and the current docs let users switch between the Opus and Sonnet models their organization allows, per thread or as a channel default, with each reply's footer naming the model that answered.
You need a Claude Team or Enterprise organization on Anthropic's own service, without Zero Data Retention or customer-managed encryption, an Owner in that organization, and a Slack workspace admin. Free, Pro and Max plans and third-party deployments are excluded. Slack users do not each need a Claude seat to use it in channels. On a Team plan, nothing runs until the usage balance is funded.