Claude in Slack for a Small Team: What Claude Tag Does, What It Costs, and When an Agent with Per-Person Access Fits Better

Your team already pays for Claude, and now Claude in Slack has turned into something new called Claude Tag. A small firm needs to know whether it can take the chasing, the Monday status update and the CRM edits off your plate, or only the writing and summarising. Get it wrong in one direction and you keep paying a coordinator for work Claude could do. Get it wrong in the other and a shared bot edits a client record or sends a message nobody in particular approved. This guide sets out what Claude in Slack does today, which plan it needs, how its shared identity works, and a simple rule for which jobs to give it and which jobs need an agent that works with each person's own access.

The short answer

1. Use Claude in Slack for reading and writing work inside the conversation: summaries, answers, file analysis and drafts. That is what Slack's own help page for the Claude app lists.
2. Expect Claude Tag to replace that app. Slack's page gives the switch date as 3 August 2026, and the new version is in beta for Claude Team and Enterprise customers, according to TechCrunch's launch report.
3. Remember that a whole channel shares one Claude. Admins choose its tools and data per channel.
4. Before it changes a deal, sends an email or posts for a client, find out whose access it uses and who approves the action. If you cannot answer both, keep those jobs with a person, or with an agent that uses each asker's own access and asks that asker to approve.

The rest of the article explains each step and gives you a table, a worked thread and a checklist.

What Claude in Slack does today

The original app is simple. You message Claude directly or mention it in a channel, and it works with the conversation in front of it. Slack's help page describes four jobs: summarising conversations, answering questions, analysing files and drafting content. In a thread it reads the earlier replies and answers in the thread, not in the main channel. It has limits worth knowing. It reads the 20 most recent messages in a channel or the 50 most recent replies in a thread, and your Slack conversations with Claude do not sync to your Claude account on the web.

It also needs a paid Claude seat. The help page lists Pro, Max, Team and Enterprise accounts, and the Slack side works on any Slack plan. On an Enterprise org, admins can install it across workspaces and choose who may use it.

What changes with Claude Tag

Claude Tag is Anthropic's replacement for that app, and the model is different. In the VentureBeat launch coverage, Anthropic describes one Claude that works with everyone in a channel, not a separate copy for each user. Anyone in the channel can see what it has done and carry the conversation on.

Four other details from the launch coverage matter to a small team:

• Admins scope it. An admin picks which tools, data and channels each Claude identity can use, and can set up separate identities for separate jobs. TechCrunch gives the example of a legal Claude that cannot add memories to the engineering channel.
• It can act without being asked. With its ambient mode on, VentureBeat reports, it raises relevant information and follows up on threads or tasks that went quiet.
• There is a log and a budget. Admins can see every action Claude took and which person asked for it, and can set token spending limits for the whole organisation and for each channel.
• It is a beta. VentureBeat also notes that Anthropic has not published detailed pricing for Claude Tag itself.

One Hacker News commenter summed the launch up as one Claude per Slack channel, and another expected that many companies will simply switch it on because they already have Slack and their staff already use Claude. A later comment reports that Claude Tag turned out to be useful for a few things. These are individual reports, not measurements, but they match the official picture: a shared assistant that knows the channel.

What it costs a small team

The seat is the visible cost. On Claude's pricing page, the Team plan is for teams of 2 to 150 and lists Claude Tag among its features. A standard seat costs $20 a month billed annually or $25 billed monthly, and a premium seat costs $100 billed annually or $125 billed monthly. Enterprise starts at 20 seats, with usage billed on top.

For a ten-person agency on standard Team seats, that is $200 to $250 a month before any Claude Tag usage, which the launch reports describe as token-based with admin spending caps. If your team already pays for Claude to write and research, the Slack side may cost little extra. If you would buy Claude seats only to get a bot in Slack, price the whole team, because the shared identity is only useful when the people in the channel can use it.

Whose access does it use?

Claude Tag gives a channel one shared Claude identity with tools an admin grants, so before it writes anything, settle whose access it acts with, because none of the public pages about it say. Slack's help page does not cover it. TechCrunch and VentureBeat say admins grant tools and data, and VentureBeat says the action log records who asked for each task, but neither says whose CRM login, mailbox or file permissions are used when Claude acts.

For reading and writing inside Slack this hardly matters. For a small firm's coordination work it matters a lot, because those jobs touch other people's systems:

• Updating a HubSpot deal changes a record that your sales lead owns.
• Sending a follow-up from someone's Gmail puts words in that person's name.
• Posting an update in a shared client channel reaches the client.

When one identity holds one set of admin-granted tools, every person in the channel borrows the same access. A junior teammate can ask for a change that only the founder should make, and the log will show who asked only after it has happened. Treat this as a recommendation drawn from what the sources leave out: until Anthropic's own documentation states how Claude Tag handles credentials and approval for writes, treat those jobs as not yet delegated.

Which jobs go where: a decision table

Use this table to sort your team's recurring work. It is built from the Slack help page, the launch reports, Claude's pricing page and Kipwise's own product notes. The Kipwise column describes a different product, an agent that runs in Slack with each person's own connections. It is not Claude and not Anthropic's app.

The table gives a simple rule. Jobs that change a record or send something in a person's name fit an agent that uses that person's own access and asks that person to approve each change. Jobs that only read and write text inside the conversation fit Claude well, and there is no reason to move them.

A worked example: chasing client requests in a thread

This example is illustrative, with a made-up agency. The ops lead at a twelve-person agency wants to know which client requests from this week still have no reply.

With Claude Tag in the client channel, she can ask it to summarise the week's thread and list the open requests. That is reading and writing inside Slack, and it is what the product is described as doing well.

The next step is the one that needs care. She wants the HubSpot deal for the slowest client marked "waiting on us" and a short follow-up drafted in her Gmail. With a shared identity, that change runs with whatever HubSpot access the admin gave the channel's Claude, and the public pages do not say how it is approved.

Here is the same thread with the Kipwise agent, using only what its product notes state. She mentions the agent once in the thread: "which client requests from this week have no reply yet?" Plain replies in the same thread continue the work, with no further mention needed. When she asks it to update the deal and draft the follow-up, it uses her own HubSpot and Gmail connections, which connect separately for each person. Each change waits for an Approve or Deny card that only she can press, and the card expires after 30 minutes. If she lets it expire, nothing runs. To get the same check every Monday, she turns it into a scheduled task, approves it once, and later runs do not ask again.

Both tools can sit in the same Slack workspace. Claude keeps the drafting and analysis; the per-person agent takes the jobs that touch other systems.

Failure modes to plan for

• Shared memory in a client channel. Claude Tag learns from the channel it follows. If client staff are in a Slack Connect channel, decide what you are comfortable having a shared identity remember before you add it.
• Ambient follow-ups that sound like you. A bot that follows up on quiet threads by itself can be helpful inside the team and awkward in front of a client. Turn the ambient behaviour on in internal channels first.
• Over-granted tools. Because admins grant tools to an identity, one generous setup gives everyone in that channel the same reach. Create a narrow identity for each channel, as the launch coverage describes, rather than one broad one.
• Context limits. The original app reads only recent messages, per Slack's help page. For a month of history, export or summarise in stages.
• Plan surprises. A Pro or Max seat covers the old app but not Claude Tag, which launched for Team and Enterprise. Check the plan before you promise it to the team.

A checklist before you turn on any shared AI identity

Run this before you add Claude Tag, or any shared assistant, to a channel where it can change something:

1. List the tools you will grant it, and remove any it does not need for this channel.
2. Write down whose account each tool uses. If you cannot find out, mark that tool read-only.
3. Decide who approves a change and how you will see it before it runs, not after.
4. Set a channel spending limit and an owner who reviews the action log each week.
5. Keep client-facing sends with a named person until steps 2 and 3 have clear answers.

How to tell it is working

Give it two weeks and track three numbers: how many summaries and drafts the team used without rewriting, how many follow-ups went out on time, and how many changes in other tools someone had to undo. If the first two rise and the third stays at zero, widen its scope one tool at a time. If changes had to be undone, move those jobs to a per-person setup with approvals. For how Slack's own built-in AI compares, see what each Slack AI plan includes and where it stops, and for a wider field, the list of AI coworkers for Slack.

What to do next

Write down your team's five most frequent coordination jobs and mark each one "text only" or "changes something". Give the text-only jobs to Claude in Slack this week. For the jobs that change something, try the Kipwise agent on its install trial, $2 for 7 days per the pricing page, with one chasing job in one channel, and have the person who owns that job press Approve on every change.