Claude Tag is a shared artificial intelligence agent living inside your Slack channels. Any member of the channel writes a message mentioning Claude, hands it a task, and Claude answers in the thread with what it produced. Anthropic introduced it in June 2026 in its launch announcement, and the product remains in public beta, checked in November 2026.
What sets it apart from a regular assistant is sharing. There is one Claude per channel, not one per person. It keeps the channel's context between conversations. A colleague can pick up a thread someone else opened the day before. For a manager, three questions matter before switching anything on. What Claude sees, what it bills, and what Anthropic keeps.
What exactly is Claude Tag?
Claude Tag turns Slack into the working surface for an AI agent. An admin installs the app and picks the channels Claude enters. They then connect the tools it may use, a code repository, an issue tracker, a data warehouse. After that, conversation is enough. Claude breaks a request into stages, runs each stage with the tools it is allowed, and posts its progress in the thread where the request was made.
The product replaced the older Claude app for Slack in August 2026. The two do not coexist. It runs on Anthropic's Opus model and stays limited to the Team and Enterprise plans. Free, Pro, Max and Ultra give no access. Anthropic claims most of its product team's code comes out of its internal version of Claude Tag. The vendor publishes that figure about itself, so it should be read as such.
What Claude Tag sees in your Slack
This is the question that comes up in every scoping workshop, and the answer reassures more than it worries. Claude reads the full history and acts only in the channels a human added it to. Elsewhere it has keyword search across public channels, exactly like any member of the workspace. It cannot invite itself into a private channel.
Credential isolation follows the same logic. A credential attached to a single channel stays invisible in every other one. The Claude Tag documentation stresses a point teams underestimate. A credential placed on a public channel becomes usable by everyone who joins that channel afterwards, including six months from now.
How to install Claude Tag in a Slack workspace
Setup takes two pairs of hands, by design. A Slack admin installs the app, which adds it to no channel at all. They then post a connect command in a channel, get a code valid for a quarter of an hour, and pass it to the owner of the Claude organisation. That Slack admin does not even need a Claude account.
- On the Claude side, the owner pairs the workspace, sets the monthly spend limit, then adds Claude to the first channels.
- On connectors, each external access applies either to all of Slack, to one workspace, or to a single channel.
- On security, credentials live outside the execution sandbox and are injected at the network boundary, so the model never sees them.
- On traceability, Claude acts under its own service accounts, which makes its actions identifiable in every tool's logs.
One piece of field advice before connecting the first tool. Create a dedicated service account per application rather than reusing a personal login. You will be able to revoke the agent's access without cutting off a colleague, and you will be able to tell who did what.
What Claude Tag costs
The commercial model is surprising, because it looks nothing like per-seat software. Claude Tag is included in the Team and Enterprise plans, and channel work requires no seat for participants. A salesperson with no Claude account can mention Claude in a channel and get an answer. Usage is drawn from an organisation balance funded by the owner, with a monthly cap the owner sets. The pricing page puts the Team seat at around 20 euros a month billed annually, and around 25 euros month to month.
Direct messages follow another logic. They run on the member's own Claude account, with their personal connectors, and count against their seat's limits. An owner can also switch direct messages off for the whole organisation. If usage-based billing is new ground for you, our token guide explains how a budget slips and how to hold it.
Ambient mode, or when Claude speaks without being asked
This is the most discussed and the most misunderstood feature. With ambient behaviour enabled, Claude flags useful information by itself, revives a stalled thread, chases a forgotten task. Many read that as a form of autonomy. The mechanics are simpler than the promise.
At Tandem we set up an open-source autonomous agent on a dedicated server. We wired it into our own Slack, before Claude Tag shipped. I wrote up the experiment in my edition on AI employees. Two lessons still hold. Proactivity rests on a periodic agent turn and on scheduled triggers, nothing magical. And the bill climbs fast as soon as the model in use thinks a lot.
My verdict at the time still stands. The value came from the interface, not from the model. Talking to an agent inside Slack changes a team's adoption rate, because nobody has one more tool to open. That is exactly Claude Tag's bet, and that is what makes it interesting for a small company. The monthly spend cap also answers the cost problem we ran into.
Should you install Claude Tag or build your own Slack agent?
The build or buy question comes up exactly as it does on other AI building blocks, and I devoted a LinkedIn comparison to it in July 2026. In 2023 everyone wanted to build. Since then off-the-shelf products move faster, and the technical debt of a homemade agent is paid every month.
For an agent inside Slack, Tandem now recommends buying. Claude Tag covers the essentials without a line of code, with an access governance model an internal team would spend weeks writing. Building still makes sense in two cases. When your organisation refuses the data retention described below. And when the agent has to run inside your own network, on your own servers.
What Anthropic keeps, and for how long
Here is the point that most often stalls a security committee, and it deserves reading before signature. The data lifecycle page states that Claude Tag keeps two things on its side, session transcripts and channel memory. During the beta no automatic retention period applies, and your organisation's retention setting does not cover these items.
| Action | Effect on Claude-side data | Does Claude still answer? |
|---|---|---|
| Remove Claude from a channel | Nothing is deleted, everything returns if re-added | No, in that channel |
| Delete the channel in Slack | Channel sessions, memory and routines deleted | Not applicable |
| Disconnect the workspace | All of that workspace's data deleted | No |
| Uninstall the Slack app | Same effect, plus the install token | No |
Three practical consequences follow. An organisation on zero data retention or under a customer-managed encryption key simply cannot switch Claude Tag on. Deleting a message in Slack does not remove it from a transcript already recorded. And there is no button to delete one thread's transcript, the finest grain remains the channel.
These trade-offs are not settled case by case in a corridor. They belong in a written framework, saying which channels get an agent and who decides. Our governance guide sets out the roles to define before opening the first access.
Is Claude Tag ready for a small or mid-sized company?
Yes, for a team already on Slack and already on a Team or Enterprise plan, provided you start small. The product is in beta, its behaviour may change, and Microsoft Teams is not served to date. If your company lives in Teams, the question comes back later. For everyone else the entry ticket is low, since no extra seat has to be bought for the participants of a channel.
Tandem supports more than 50 companies on this kind of rollout, and the same pitfall keeps coming back. Teams open too many channels at once, nobody knows what the agent sees any more, and the security committee shuts it all down three weeks later. One channel, one repetitive use case, one spend cap, then widen.
Two ways to go further. To give Claude your internal procedures rather than re-explaining them every week, read our skills guide. To run a recurring task that posts its result into a channel, Claude routines cover the need. And if the real issue is getting your teams to actually use it, our Claude training starts from your own use cases.
