Cloud Tag for Claude Tag users
Cloud Tag is built to work like Claude Tag, in email instead of Slack. If you know Claude Tag, almost everything carries over: you mention the agent in a thread, it works in a sandbox, and admins configure it per channel. This page lists what maps one to one, what's different, and what isn't built yet.
A Cloud Tag page that has a Claude Tag counterpart keeps its path. For claude.com/docs/claude-tag/users/commands, open /docs/users/commands on your Cloud Tag site.
What maps one to one
| Claude Tag | Cloud Tag |
|---|---|
| Slack workspace | A linked Gmail mailbox (an agent), in an organization |
| Channel | Google Group |
| Thread | Email thread |
@Claude |
@cloud, or whatever you name the agent |
| Session per thread; steer by replying | The same |
| Respond automatically | The same, per group |
| Channel, workspace and org instructions | Organization, agent, bundle and group instructions, added in that order |
| Model per channel, inherited | Model and effort per group, agent and organization |
| Channel memory | Group memory, on the Memory page |
| Workspace memory | An agent's shared notes, written only from groups that allow it |
| Connectors with shared credentials | MCP connectors (none, bearer token or OAuth), credentials in Anthropic Vault |
| Org-scope connectors, skills and credentials | Organization-wide connectors, skills, secrets and repositories |
| Bundles and scope rules | Bundles, applied by group address pattern, by agent, or by hand. The narrowest scope wins for one server or repository |
| Personal connectors | Your connectors: your own accounts, used only for your requests; read-only tools allowed per thread, and every other call approved by you from an emailed link |
| Google Drive connector | Drive as the agent's own account, scoped to the files each group can reach |
| GitHub repositories | Repositories cloned into each session, with the agent's GitHub account or a token you provide |
| Auto mode allow rules | Auto-approve rules per organization and group |
| Network audit | Network activity on the group's Network & secrets page |
| Credentials for allowed hosts | Secrets, sent only to the hosts you list |
| Tool permissions and approvals | Presets and per-tool overrides; approvals on the web |
| Plugins and skills | Built-in document skills and uploaded custom skills. Uploading a plugin .zip adds each of its skills; its commands, agents, hooks and MCP servers are not used |
| Allowed websites | Network mode: full, limited (allowlist) or none |
| Spend limit: org, default per channel, per channel | The same three tiers, with emails at 75%, 90% and 100% |
| Routines (cron, time zone, off after 3 failures) | The same; each run starts a new thread |
@Claude connect pairing |
Add the mailbox in Google Groups, then answer Yes. @cloud connect sends the same ask |
!help !configure !mute !unmute !restart !status !fast !feedback !routines !fork |
The same commands, as the first line of an email. !routines #group and !fork #group name another group you post in |
!status seen only by you |
Emailed only to you |
| Reply footer with model and Configure | The same |
| Owner, Claude Tag admin, channel manager | Organization Owner, Admin, group admin |
| Activity page | Each group's Usage & activity page |
What's different
- Billing. There are no seats or usage balance: each organization brings its own Anthropic API key, and everything is billed to that Anthropic account.
- No top-level messages. Every email is in a thread, so every conversation is a thread session.
- Approvals happen on the web. Email has no buttons, so the agent posts a link to the Approvals page.
- Everyone who signs in gets an organization. The first sign-in creates an empty one, even for someone who only manages a group.
- No live progress checklist. The agent replies when it's done. Use
!statusto check on it. - No emoji reactions or edits. Email has neither. A message deleted in Google Groups is still in the agent's archive.
- History before it joined. Google Groups has no API for its archive, so earlier mail comes in by mbox import instead of backfill.
- Which groups it can join. A Workspace mailbox joins only its own domain's groups; a personal Gmail mailbox only public googlegroups.com groups. Each agent in a group, including other organizations' agents, serves it separately.
- Who can use it follows the group's posting permissions in Google Groups.
- Drive is scoped per group. Each group reaches only the Drive files linked in its mail, shared with its address, or shared with its whole Workspace domain.
- GitHub uses the agent's own GitHub account, which an admin signs in as once, or a token you create, not the Claude GitHub App.
- Secrets can't override each other. A variable name can be set in only one place for a group.
Not built yet
- A sensitive-content hold for personal connector results
- Whole-Drive access for a group
- Group DMs (one-to-one email DMs are available)
- Asking for a different model in a thread (
!fastis the per-thread switch) - Channel watching and pull request subscriptions
- GitLab access, and federated identity
- Auto mode as a whole, beyond the allow rules
- A sandbox-level network log
- Restricting use to members of your organization or to roles
- "Managed by" instructions from another group
- Uploads above 4 MB
- Channels other than Google Groups, such as Google Chat or Slack
Moving a channel over
- Set up Cloud Tag and add the agent to the Google Group that replaces the channel.
- Copy the channel's instructions into the group's Custom instructions, and your org and workspace instructions to the organization and agent.
- Add the same connectors on the group's Connectors page (the organization's for org-scope ones, or a bundle for workspace and pattern scopes), signing in with the same service accounts. Re-add credentials as secrets. Re-add repositories with the agent's connected GitHub account, or a token.
- Copy auto mode allow rules into the organization page or the group's Approvals page.
- Turn on the same skills, and upload custom ones. A plugin zip you uploaded to Claude Tag adds all of its skills in one upload.
- Recreate routines on the Routines page.
- Set the same spend limits.
- Ask members to reconnect their personal connectors on Your connectors.
- Ask the agent to remember what channel memory held, or upload it as a group file.