# Claude Tag in Slack, and the open-source alternative you can audit

> Claude Tag is Anthropic's Slack agent: tag @Claude in a channel and it works under your organization's identity with the tools an owner granted. Kortix is the open-source AI Operating System, and its Slack channel puts a named sender on every message, with a reach an admin can read.

URL: https://claudetagalternative.com/claude-tag-slack

## The identity link and the join policy, as Kortix documents them

`channels`

```
A Slack or Teams message carries no Kortix credential.
Each person links a Kortix account once.
After that, Kortix knows who sent each message, and the session runs on behalf of that person.
The join policy decides who may join a session that a channel message started.
Each binding has one policy. New sessions use it. An existing session keeps the policy it started with.
```

Source: https://kortix.com/docs/connect/channels

A message carries no credential, every sender links an account once, and each binding carries one policy.

## What Claude Tag is inside Slack

Claude Tag is Anthropic's Slack agent. Tag @Claude in a channel and it takes the task. It works through the job in stages, then answers in the thread with what it produced. It reads the channel it sits in and keeps what it learns, so the next request starts with context behind it. Anthropic gives it three surfaces: channel mentions, direct messages, and a panel that opens from Slack's AI assistant header ([Anthropic's help article](https://support.claude.com/en/articles/15594475-what-is-claude-tag)).

The decision that matters for an admin sits underneath the tagging. Claude Tag is a member of the workspace with an identity of its own. In a channel it acts under your organization's identity, with the tools and access an owner set for that channel. The work is billed to the organization. A direct message runs on the person's own Claude account instead, with the tools they turned on there.

Two agents share one app, with different owners, budgets and auditors. Most of the admin questions follow from that split.

Claude Tag replaced the earlier Claude in Slack app on August 3, 2026, and administrators opt in to the move within 30 days ([Anthropic](https://support.claude.com/en/articles/15594475-what-is-claude-tag)). Slack's help page walks through the app install and points readers to Claude Tag for how it now works ([Slack](https://slack.com/help/articles/53532192117267-Use-Claude-in-Slack)).

## Only an owner can switch it on

Setup is not a personal decision, and it is not the admin's either. After the Claude app is installed, a Primary Owner or Owner provisions Claude's identity. That person connects the organization's tools and repositories, then chooses the channels Claude Tag may work in. Anthropic states that the Admin role cannot do this ([Anthropic](https://support.claude.com/en/articles/15594475-what-is-claude-tag)).

Slack's side is a separate permission system. Any member with permission to install apps can add Claude. In an Enterprise organization an org owner or Org Admin installs at the org level and picks the workspaces that get it. Then the app has to be added to a channel before anyone can tag it.

Put the two together and the first question for your team is who holds both keys. If the Slack install sits with an integration admin and the Claude identity sits with a Primary Owner, one person can widen the other's decision by adding a channel. Write down both names.

## What it can reach, and who can prove it later

Reach is granted in three layers, each inheriting from the one above: organization-wide credentials and repositories, then workspace access for every public channel, then extra credentials on a private channel. A legal channel can hold its tools and memory apart from an engineering channel.

Channels with Slack guests get their own setting. Anthropic's three choices are Restrict, the default that keeps Claude out of those channels; Channel only, where Claude answers with channel-only access while a guest is present; and Allow, with the full configured access ([Anthropic](https://support.claude.com/en/articles/15594475-what-is-claude-tag)).

Attribution is where to spend the audit hour. Anthropic says every action can be traced in the tool where it happened. Slack posts come from the Claude app, and commits and pull requests carry the Claude GitHub App as author, with a link back to the thread that started them. The Activity page under Organization settings lists every scheduled and one-time task, and every network call made with the organization's agent identity. Only Owners can open it.

Take one merge from last month and read its author. When that author is the Claude GitHub App, the commit links back to the thread that asked for it, and the Activity page holds the network calls made under the organization's identity.

## What it costs, and where the meter stops

Claude Tag is consumption-based, so spend follows usage and not headcount. A Primary Owner or Owner sets a hard cap for the whole organization, per-channel limits on top of it, and alerts at 75% and 95% of any limit. Work that would cross a limit is declined, not quietly cut short. The blocked person can ask an admin for more from inside Slack ([Anthropic](https://support.claude.com/en/articles/15594475-what-is-claude-tag)).

The bill is one number with visible ceilings and a per-channel breakdown. What it does not give you is the cost of a finished task: a question about a document and a job that runs for hours draw on the same cap.

## Where Claude Tag stops

Claude Tag is in beta on the Team and Enterprise plans as of August 2026. Beta status is Anthropic's own word, and it covers the parts an admin leans on: the setup screens, the spend controls, the identity model. Plan for them to move.

It works in Slack. Anthropic frames Slack as where the product starts, and no other surface carries a committed date.

There is one shared Claude per channel. That is the feature, since anyone can steer it or pick up a task someone left. It is also a single queue: two unrelated requests in one channel share one memory and one working context.

The model is Anthropic's, chosen by Anthropic. A model change on their side reaches your agent without a change request on yours, which is a dependency to record for any process with an owner.

## When Claude Tag is the better pick

For a lot of companies, Claude Tag is the better pick. If your team already pays for Claude Team or Enterprise, the work is conversational: summaries, drafts, ticket triage, a first pass on a document. Nobody needs the output to arrive in a repository under review. In that case Claude Tag is less setup than anything you host. No server, no repository, no model key, no connector catalog. The install is a Slack app and an owner's afternoon.

## Where the reach and the audit trail live

Kortix is the open-source AI Operating System, and its Slack channel answers the same admin questions from the other side. A Slack message carries no credential, so each person links a Kortix account once with `/kortix login`, and that link is what puts a name on the action. Kortix checks that the sender may run the agent, then runs the session on that person's behalf. The grant belongs to the agent: what it can call does not widen or narrow with the sender's role, and the sender's own secrets and provider keys reach only a DM session, private to that person. [Kortix](https://kortix.com/docs/connect/channels) documents each of those steps.

Join policy is the second lever, and it governs who may enter a session that a channel message started. `project_open` lets any linked member in. `owner_approval` gives the session owner an Approve and a Deny for each new person. `owner_only` keeps the session to the person who started it. A binding carries one policy, and a session keeps the policy it started with.

Reach is declared in the project instead of clicked in a vendor console: "3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP." Connectors, secrets and grants are files in a repository you can read, diff and review, and each session gets its own machine: "One isolated sandbox per session. Each session has its own isolated machine and branch."

Slack and Microsoft Teams are live. Email is experimental, per project.

Attribution lands where a reviewer already works. "Session work reaches `main` through a change request. Merge is default-deny for agents." A review card posts into the thread, and the merge waits for a person holding `project.gitops.merge`. A merged change request on Kortix's own repository shows the review a person reads: what changed, which endpoints are affected, a risk note, and the tests that ran ([a merged pull request](https://github.com/kortix-ai/suna/pull/10819)).

The trade runs in both directions. Claude Tag gives you an agent in Slack this afternoon, with Anthropic holding the identity, the model and the memory. Kortix puts the identity, the reach and the merge in a repository your company owns, where an auditor can read them.

## Claude Tag questions owners ask

### What is Claude Tag?

Claude Tag is Anthropic's Slack agent: you tag @Claude in a channel and it takes on the task with the tools and access an owner granted that channel, working under your organization's identity. It also answers in direct messages and in Slack's AI assistant panel.

### Is the Claude Tag available now?

Yes, in beta on the Claude Team and Enterprise plans as of August 2026. Claude Tag replaced the earlier Claude in Slack app on August 3, 2026, and administrators opt in within 30 days to move over.

### Does Claude Tag cost extra?

Yes. Channel work is consumption-based and billed to your organization, and direct messages are billed to each person's own Claude account. Owners set an organization cap and per-channel limits, and work that would cross a limit is declined.

## Run your first Slack session on a machine you own

Connect Slack, link your account, and let the first change request land in the thread.

[Try Kortix](https://kortix.com) · [Read the docs](https://kortix.com/docs/connect/channels)