Anthropic's Claude Tag API lets IT manage Claude in Slack from code

Anthropic added a Claude Tag API to its platform on 23 October 2026, according to the company's release notes. It lets a Claude Enterprise organization read and change, from its own systems, the settings that an owner otherwise sets by hand in claude.ai. Claude Tag is Claude working as a teammate in Slack: someone mentions Claude in a channel and it answers in the thread. The API is in beta and is not available to Claude Console organizations.

What you can do with it

Anthropic's documentation names three uses: turning Claude on channel by channel from an approval workflow, keeping configuration in version control, and auditing what applies in every channel. Its 14 endpoints cover setting up, listing, reading, changing and removing Slack channels; listing, reading and changing Slack workspaces; listing Access bundles; and attaching or detaching a bundle to a channel or workspace.

Anthropic's description of the limits is just as specific. This release covers Slack only. Connecting a Slack workspace, creating or editing bundles, spend limits and every other Claude Tag setting stay in claude.ai.

The five settings, and what you get by default

• enabled: whether Claude is on. Anthropic says the nearest level that sets it wins, so true on a channel turns Claude on there even if its workspace stores false. If no level sets it, the value is true.
• instructions: standing guidance of up to 8,192 characters. Unlike the others, it is additive: the text from every level is joined, broadest first.
• model: the model new sessions start on. If no level sets it, Claude uses the product's default model.
• environment_id: the cloud environment sessions run in. It is checked when written and not again, so a later change to the environment is not reflected in the stored value.
• guest_policy: allow, restrict or channel_only, applied to everyone in a channel while a Slack guest is present. If no level sets it, the value is restrict.

Two of those defaults matter for a rollout. According to the documentation, a channel that nobody configures resolves to Claude on, so an approval process that assumes silence means off should check what its workspace and organization defaults store. The guest default points the other way: Claude does not answer in a channel with a guest in it unless someone chooses otherwise.

Auditing with include_inherited

Settings exist at three levels: the channel, its workspace, and the organization's Slack defaults, which only an owner can change in claude.ai. A plain read returns what a level stores, with null for anything unset. Adding include_inherited=true to a read returns what applies once all levels are combined, plus a sources field naming the level each value comes from. Anthropic's example shows a channel with Claude turned off by its workspace, a model set on the channel itself and instructions inherited from the Slack defaults.

Anthropic also lists source types that complicate an audit: a parent channel for channels started from another channel, and a channel shared across a Slack Enterprise Grid, where an enabled value stored under another workspace of the Grid can apply. A script that only compares a channel's own record with a spreadsheet will miss both.

What the documentation leaves open

• The notes give no date for general availability, and the beta header name implies the interface may change.
• A channel or workspace your organization does not own returns the same 404 as one that does not exist, so a missing channel and a permissions problem look identical.
• An Admin API key belongs to the organization and keeps working after its creator leaves; the documentation tells teams not to share it or commit it to source control.

It also does not say what Claude Tag costs or how much a channel uses; spend limits stay in claude.ai, and the Usage and Cost Analytics API released on 29 October covers API and Enterprise usage in separate documentation. Whether Claude Tag usage appears in it is not stated on the page I read.