How Do You Use Claude Code With Slack?

How Do You Use Claude Code With Slack?

You mention @Claude in a channel with a coding task, and Claude Code picks it up in Anthropic's cloud, works on your repository, and posts a pull request link back into the thread.

That is the simple version. The complicated version is that two different products now sit behind that same @Claude handle, one was retired in channels on October 5, 2026, and which one answers depends on whether anyone in your organization has paired your Slack workspace. Get that wrong and you will follow a setup guide to its end and get a retirement notice instead of a reply.

This guide separates the two, gives the current setup path, and names what Anthropic's documentation leaves unspecified.

What is the short answer?

It depends on who you are.

• On Team or Enterprise, the answer is Claude Tag. An Owner pairs the workspace once, and everyone in a channel can then use @Claude with no per-user setup. Code work comes back as a draft pull request opened under the Claude GitHub App.
• On Pro or Max in an unpaired workspace, the answer is the older Claude Code in Slack. You connect your own account, authenticate your own repositories, and sessions run against your personal plan limits.
• If your workspace is paired and a channel is still set to Legacy, nothing works. Anthropic's notice reads: the legacy Claude in Slack bot is retired effective October 5, 2026 and no longer responds in channels.

Both versions share one thing: neither is a bridge to a session on your laptop. Both run Claude Code in Anthropic's cloud against a repository you have connected.

Automating work is only worth it if the foundations underneath are sound. Get a free SEO audit and find out what is actually holding your site back.

What exactly is Claude Code in Slack?

Anthropic describes it as intent detection bolted onto an existing Slack app: mention @Claude with a coding task, and Claude detects the intent and creates a Claude Code cloud session.

The important detail is that it is not a separate app. Anthropic states it plainly — the earlier Claude in Slack and Claude Tag are two versions of the same @Claude Slack app, not two apps, so there is nothing to uninstall. Admins switch a workspace or an individual channel between Legacy and New.

The legacy per-user version

Each user links their own claude.ai account. Sessions spawn per request, there is no memory between them, and pull requests open under the asking user's name. That last detail is Anthropic's own diagnostic test: if @Claude still opens pull requests under the asker's name, that channel is on the legacy version.

It requires Pro or Max, a workspace not connected to Claude Tag, cloud sessions enabled, and a GitHub account connected at claude.ai/code with at least one repository authenticated. GitHub only — GitLab is not supported.

Claude Tag, the current version

Claude Tag runs @Claude as your organization's shared identity with admin-configured access. The differences are structural: one agent identity with org-level service credentials instead of per-user links, one persistent session per thread shared with the channel, per-channel memory instead of none, and standing work through routines and channel watching.

It also changes the billing shape. Adding Claude to Slack does not add a per-seat charge, and Slack users do not each need a seat to work with Claude in channels. Usage draws against an org balance capped by an Owner-set monthly spend limit.

The two things people confuse it with

Two other integrations share the word Slack and do something else entirely.

• The Slack connector for claude.ai points the other way, letting Claude on claude.ai search your workspace's channels, DMs and files for context. Team and Enterprise only, and it requires the Slack app installed first.
• A Slack MCP server points the other way again, giving Claude Code in your terminal tools to read and write Slack through Slack's own endpoint at https://mcp.slack.com/mcp. The developer installs it themselves.

How do you set it up?

The Claude Tag path

Setup requires the Owner role in a Claude organization on a Team or Enterprise plan, plus a Slack workspace admin, because only a workspace admin can run the connect command.

1. Check the prerequisites. Team or Enterprise on Anthropic's first-party service, no zero-data-retention, customer-managed-key or HIPAA configuration, at least one public channel, and routines enabled.
2. Install the Claude app from the Slack Marketplace.
3. Invite and connect. Run /invite @Claude in a channel, then send @Claude connect as a standalone message. Claude replies with a pairing code that works once and expires in 15 minutes.
4. Paste the code into your Claude admin settings and pair the workspace.
5. Connect GitHub if you want code work. This installs the Claude GitHub App, and only an owner of your GitHub organization can install it.
6. Set a spend limit and launch. Pick the channels Claude joins and notify members.

One prerequisite fails in a confusing way. Until routines are enabled, Claude answers every mention with a reply saying it is unavailable, and does no work. If @Claude responds but never does anything, check that first.

The legacy path

A workspace administrator installs the Claude app from the Slack Marketplace. Each user then opens the app's Home tab, clicks Connect, and completes the browser auth flow, then configures cloud sessions at claude.ai/code and authenticates at least one repository.

Then pick a routing mode in App Home: Code only sends every mention to Claude Code, or Code + Chat lets Claude route each message. Finally invite Claude with /invite @Claude, because it joins no channel automatically.

What about /install-slack-app?

It is real, and documented in exactly one place: the Claude Code CLI commands reference, where it sits beside /install-github-app and is described as installing the Claude Slack app by opening a browser to complete the OAuth flow. It is not mentioned on the Slack setup page at all, which sends you to the Marketplace instead. Two Anthropic pages, two install routes, no cross-reference.

Wondering whether your own stack has gaps like this one? Run a free audit and get a clear picture before you build on top of it.

How do you actually use it in a channel?

In the legacy version the flow is fixed and visible. You mention Claude, it creates a session on claude.ai/code, posts status updates to the thread as work progresses, then mentions you with a summary and action buttons: View Session, Create PR, Retry as Code, and Change Repo. It works in public and private channels but not in DMs.

Claude Tag is looser. An at-mention guarantees a response, but Claude may also reply to a message that does not mention it when it judges a reply is warranted, and once a thread is active it follows replies there. It works in channels, group DMs and one-to-one DMs.

Two Claude Tag behaviors surprise people. Anyone in the channel can steer a running session by replying in its thread, not just whoever started it. And the progress checklist is edited in place — which matters, because Slack sends no notification when a message is edited, so a thread can look frozen while the list is still moving.

It also understands a small set of bang commands sent with a mention, including !help, !status, !restart, !mute and !routines.

When should you use each one?

• Use Claude Tag when more than one person needs to ask, when you want standing work like a recurring digest, or when you do not want to buy everyone a seat.
• Use the legacy version only when your workspace is unpaired and you have Pro or Max. It is an individual's fallback, not an organizational choice — you cannot switch a workspace or channel back to Legacy anymore.
• Use a Slack MCP server when the work is in your terminal and Claude Code needs Slack as a tool.
• Use the Slack connector when the work is in claude.ai and Claude needs Slack for context.

Picking tools is easier when you know where the real bottleneck is. Start with a free audit and get an outside read on what to fix first.

What are the limitations worth knowing?

The prompt injection warning

Read this one twice, in Anthropic's own words: when @Claude is invoked in Slack, Claude is given access to the conversation context, may follow directions from other messages in that context, and so should only be used in trusted Slack conversations.

Claude Tag narrows this by treating what other people post in a thread as information rather than instructions, but the underlying exposure is real in both versions.

Shared credentials are genuinely shared

Anthropic is blunt: anyone in one of those channels can ask Claude to act with the credential, so whatever the connected account can read or write is available to every member. A connector added for one person's convenience becomes a channel-wide capability.

It can search channels it was never added to

Without any connector, Claude reads its own thread and the channel's history, and workspace search can still find messages by keyword from public channels it is not a member of — the same search any Slack user has. No setting turns workspace search off.

Context is a window, not the whole thread

Mentioning @Claude partway into an existing thread gives it a window of the thread's messages, not the whole thread, with other bots' replies filtered out. In long threads, restate anything critical.

Unattended runs are not more restricted

Worth stating because most people assume the opposite. Every channel request, whether typed by a person or triggered by a schedule, follows the same path, and a recurring digest gets the same access a typed request would.

Separately, each legacy session can create only one pull request. Branch pushing and issue commenting are documented for Claude Tag but not for the legacy version.

Good automation needs good inputs. Book a free audit and see where your site is losing ground before you scale anything on top of it.

What Anthropic does not specify

The gaps here are as informative as the features.

• Slack OAuth scopes are never enumerated. No page lists what the @Claude app actually requests. Setup says only to approve the permissions.
• Session thresholds are explicitly disclaimed. Anthropic gives approximate figures for idle timeout, session age and message volume, then says the thresholds are defaults, can change, and should be treated as approximate.
• No concurrency limit and no session timeout are published for either version.
• Retention is open-ended during the beta. There is no automatic retention period for Claude Tag data, your organization's custom retention setting does not apply to its transcripts or memory, and deleting messages in Slack does not remove them from the transcript.
• Compliance tooling does not cover it yet. During the beta, Claude Tag transcripts are not included in organization data exports, and the Compliance API does not list or delete those sessions.

Anthropic also publishes an unusually candid list of controls that do not exist: no renaming the app, no per-user spend caps on channel work, no per-channel responder allowlist, no web search toggle for channels, no switch to turn workspace search off, and no enforcement of your organization's Slack session-length policy on this surface.

How does it compare to the alternatives?

The GitHub App route is the closest substitute. Mentioning Claude on an issue or pull request keeps the work where the code review already happens, which suits teams whose discussions live in GitHub rather than Slack.

The terminal remains the fastest and least constrained option for a single developer. No pairing, no spend limit, no sandbox egress rules. Slack buys you delegation and visibility, not speed.

Scheduled work is where Claude Tag pulls ahead of both, since routines and channel watching have no equivalent in the legacy version or in a plain terminal session. We covered how that scheduling layer behaves, and where its own documentation contradicts itself, in our guide to Claude Code routines.

Final thoughts

The useful thing to know about using Claude Code with Slack in 2026 is that the obvious answer is now the wrong one. Most guides still describe the per-user flow, which is retired in any workspace an organization has paired and cannot be selected again.

So establish which version you are on before troubleshooting anything. If pull requests arrive under the asker's name, you are on the legacy version. If @Claude replies that it is unavailable, routines are probably not enabled. If it returns a retirement notice, someone paired the workspace and the channel was never switched over.

Then treat the Slack surface as what it is: a delegation layer with a wide context window and a shared credential model, running in Anthropic's cloud rather than on your machine. That is a reasonable trade for team visibility, as long as you only invoke it in conversations you would be comfortable having it read.