There are four ways to get Claude into Slack, and picking the wrong one costs you a week. Claude Tag is Anthropic's Slack-native agent: it lives in your channels, keeps memory across days, and is billed by consumption rather than by seat - but it is in public beta and only available on Claude Team and Enterprise plans. Claude Code in Slack is the path on Pro and Max, works in channels only, and is built around GitHub pull requests. The Slack connector runs the opposite direction: it lets Claude on claude.ai read your Slack, and it cannot post back. Failing all of those, you can build your own bot on the Anthropic Messages API with Slack's Bolt framework, which gives you full control in exchange for owning the code forever.
Setup guideGet Claude answering in your Slack.
Get Claude answering in your Slack.
Four real paths, honestly priced.
Anthropic ships more than one Slack product and they do genuinely different things - one of them does not even run in Slack. Here is what each does, who has to approve it, and what it costs before you commit.
The short answer.
| At a glance | |
|---|---|
| Is there an official app? | Yes. Claude Tag is Anthropic's Slack-native agent, in public beta since June 2026, and it replaces the older Claude in Slack app for Team and Enterprise workspaces. |
| Plans it needs | Claude Tag: Claude Team or Enterprise only. Claude Code in Slack: Pro, Max, Team or Enterprise. The Slack connector: Team or Enterprise. |
| Who has to approve it | A Slack workspace admin installs the app. For Claude Tag, only a Slack Primary Owner or Owner can configure access and channels - a workspace Admin cannot. |
| How it is billed | Claude Tag is consumption-billed, not per seat. Channel usage draws on the organisation's balance; DMs bill the individual's own Claude account. |
| Time to set up | Minutes for the first-party apps, once the approvals land. A day or more of engineering for a bot you build yourself. |
| Biggest limitation | Claude Tag's access follows the channel it is configured in, not the permissions of the person who tagged it. Decide the channel list before you switch it on. |
| Build-your-own stack | Slack's Bolt framework plus the Anthropic Messages API. Acknowledge Slack inside 3 seconds, then do the model call asynchronously. |
Last reviewed 2026-07-29 View as Markdown
Your options.
Anthropic's Slack-related products are easy to confuse: two of them answer to @Claude and one of them runs in the opposite direction entirely. Here is the honest map.
| Approach | What you get | Where it stops |
|---|---|---|
| Claude Tag | Anthropic's Slack-native agent. Tag it in a channel, DM it, or open its assistant panel. It keeps memory across days, schedules its own follow-ups, and acts under the organisation's identity rather than yours. | Public beta, Slack only, and Claude Team or Enterprise only - not Free, Pro or Max. Access is set per channel rather than per person, so it sees what the channel sees. |
| Claude Code in Slack | The path on Pro and Max. Mention @Claude in a channel and it picks up a coding task, works on the repo and opens a pull request. | Channels only - it does not work in DMs. GitHub only, one pull request per session, and it runs on the individual's own Claude rate limits. Anthropic is retiring it for Team and Enterprise in favour of Claude Tag. |
| The Slack connector | The other direction: Claude on claude.ai searches your Slack channels, DMs and files for context when you ask it something there. | Read-only. It cannot post a message, reply in a thread or take any action inside Slack. It makes Slack a source for Claude, not a place Claude lives. |
| A bot you build | Slack's Bolt framework plus the Anthropic Messages API. Entirely yours: your prompt, your tools, your data, your model choice, and it works on any Slack plan. | You own the acknowledgement timing, the retry handling, the message chunking, the hosting, the secrets and the token bill. Budget for the maintenance, not just the build. |
| An AI teammate that already lives there | A product built for this shape from the start: it joins your channels, connects the rest of your stack, and runs whole jobs rather than answering one question at a time. Beagle is one of these. | You are adopting someone else's product decisions instead of writing your own. The only fair test is a real job in the first week. |
Step by step.
For most teams the fastest useful path is Claude Tag, so that is the walkthrough. If you are on Pro or Max, swap steps two and three for Claude Code in Slack - the surrounding steps are the same.
01 Check which Claude plan you are on
Claude Tag needs Claude Team or Enterprise. Free, Pro and Max cannot turn it on, and this is the single most common reason a setup dies at step one. If you are on Pro or Max, use Claude Code in Slack instead - it is the supported path for individual plans.
Check firstYour organisation's plan in claude.ai settings
02 Get the Slack app installed by an admin
Anthropic's Claude app has to be installed into the workspace by a Slack admin before anyone can use it. On paid Slack plans that is an approval request rather than a self-serve install, so start it early - the approval, not the setup, is usually the long pole.
WhereSlack Marketplace, then workspace admin approval
03 Have an Owner configure Claude Tag
Only a Slack Primary Owner or Owner can configure Claude Tag access and channels. A workspace Admin cannot, though an Admin can view the audit logs afterwards. It is a one-time setup in Claude's admin settings.
Whereclaude.ai/admin-settings/claude-tag
04 Choose the channels deliberately
Claude Tag's access follows the channel configuration rather than the permissions of whoever tagged it. Whatever a configured channel can see, Claude can see, for everyone who can tag it there. Start with two or three low-sensitivity channels and widen once you trust the behaviour.
05 Set spend limits before you invite people in
Billing is by consumption, not by seat: channel usage draws down the organisation's balance and DMs bill the individual's own Claude account. Set the org-wide cap and any per-channel limits up front, and note that work which would breach a limit is refused outright rather than quietly truncated.
AlertsThreshold warnings fire at 75% and 95% of the cap
06 Invite it and give it a real job
Add Claude to the channel, tag it, and hand it something you genuinely need this week - a long thread summarised, a decision tracked down, a draft written from what is in the channel. A demo question tells you nothing. A real job tells you everything.
07 Decide your prompt-injection posture before widening access
Anthropic's own guidance is that Claude may follow directions it finds in other messages in the conversation, so it should only be used in Slack conversations you trust. Before you add it to channels with external guests, shared connections or forwarded customer mail, decide how you feel about that.
What trips people up.
The things that turn a ten-minute setup into a three-week one, in the order teams usually hit them.
The plan gate, not the install
Nearly every stalled rollout is a plan problem dressed up as a technical one. Claude Tag simply is not available on Free, Pro or Max, and no amount of Slack admin work changes that. Confirm the Claude plan before you schedule anything else.
Two products answer to @Claude
Claude Tag and Claude Code in Slack both respond to the same handle but behave differently - one is a general agent, the other is a coding agent that only works in channels and only opens GitHub pull requests. If a teammate says "Claude ignored my DM", this is usually why.
The connector is not a bot
The Slack connector is the most commonly misread product in this space. It lets Claude on claude.ai read Slack. It does not put Claude in Slack, and it cannot post, reply or react. If your goal is an answer in the thread, the connector is not the thing you want.
Owner, not admin
Slack draws a line between Owner and Admin that most companies do not think about until this moment. Claude Tag configuration needs a Primary Owner or Owner. If your IT admin is not one, the setup blocks until someone senior enough sits down for five minutes.
The three-second rule, if you build it yourself
Slack expects an HTTP 200 within three seconds and retries three times when it does not get one - almost immediately, then after a minute, then after five. A model call takes longer than that every single time, so if you acknowledge the event after calling the API, your users get the same answer three times. Acknowledge first, do the work asynchronously, and dedupe on the event ID.
Slack's message limits will bite an LLM
Slack works best under 4,000 characters, truncates beyond 40,000, caps a message at 50 blocks, and accepts roughly one message per second per channel. Model output routinely blows past all of that, so plan to chunk into a thread rather than posting a wall and hoping.
Claude in Slack is still a conversation you have to start. The work most teams actually want is the job nobody remembered to ask for.
Or skip the build.
If what you want is not "Claude, but in Slack" so much as "the work done, in Slack", that is a different shape of product. Beagle joins your channels the same way, connects the rest of your stack, and runs whole jobs instead of answering one question at a time.
No plan gate to clear first
Beagle installs into any Slack workspace and starts free with 1,000 credits and no card. There is no Team-or-Enterprise tier to buy before you are allowed to try it on something real.
It reads the rest of the stack, not just the channel
Ask about an account and Beagle pulls the contract from the CRM, the thread from the inbox and the ticket from the tracker, then answers in the thread with the records it used, linked.
It starts work you did not ask for
The weekly report drafted before the standup, the renewal flagged while there is still time, the quiet account surfaced. Tagging is optional - noticing is the point.
Every outbound send waits for a nod
Beagle drafts the reply, the report and the update, and holds all of it for a human's approval. Nothing reaches a client on the AI's own authority.
Plugged into your stack.
OAuth in, every read scoped to the teammate who asked.
Common questions.
Can Claude read and reply to messages in Slack?
Yes, through Claude Tag or Claude Code in Slack. Claude Tag is Anthropic's Slack-native agent on Team and Enterprise plans: you tag it in a channel or DM it and it replies in the thread. Claude Code in Slack is the equivalent on Pro and Max, but it only works in channels and is built around GitHub pull requests. The Slack connector is a separate product that reads Slack for Claude on claude.ai and cannot reply in Slack at all.
Which Claude plan do I need to use Claude in Slack?
It depends which product you mean. Claude Tag requires Claude Team or Enterprise and is not available on Free, Pro or Max. Claude Code in Slack works on Pro, Max, Team and Enterprise. The Slack connector requires Team or Enterprise. A bot you build yourself needs only an Anthropic API key and works regardless of which Claude subscription anyone has.
How much does Claude Tag cost in Slack?
Claude Tag is not priced per seat. Usage in channels is consumption-billed against your organisation's Claude balance, while DMs are billed to the individual's own Claude account. You can set an organisation-wide cap and per-channel limits, and you get threshold alerts at 75% and 95%. Work that would exceed a limit is refused rather than truncated. You still need a Claude Team or Enterprise plan underneath.
What is the difference between Claude Tag and the Slack connector?
They run in opposite directions. Claude Tag puts Claude inside Slack, where it reads channels and replies in threads. The Slack connector puts Slack inside Claude: when you are chatting on claude.ai, it can search your Slack messages and files for context. The connector is read-only and cannot post anything to Slack.
Can I build my own Claude bot for Slack?
Yes, and it is a well-trodden path. Use Slack's Bolt framework with the Anthropic Messages API. The two things that catch people out are Slack's three-second acknowledgement window, which every model call will exceed unless you ack first and work asynchronously, and Slack's practical 4,000-character message limit, which LLM output routinely overshoots. Socket Mode avoids needing a public HTTPS endpoint, which is handy for an internal tool.
Is it safe to add Claude to a Slack channel with client data?
Treat it as a deliberate decision rather than a default. Claude Tag's access follows the channel configuration rather than the permissions of the person who tagged it, so anyone who can tag it in a channel can reach everything that channel holds. Anthropic also warns that Claude may follow instructions embedded in other messages in the conversation, which matters in channels with external guests. Start on low-sensitivity channels, set spend caps, and widen once you have watched it behave.
