Claude Tag is Anthropic's Slack product, launched on 23 June 2026, that puts one shared @Claude inside a channel with memory scoped to that channel, tasks it can schedule for itself, an admin audit log and per-channel token spend limits. It runs on Anthropic's infrastructure, works only in Slack, and runs only Claude. ChannelGate is a self-hosted alternative that runs in your own Linux environment and lets each thread pick Claude Code or Codex, along with the model and reasoning effort.

That is the short version. The rest of this page is the long one, because the choice between them rarely comes down to a feature.

If you have been looking for a way to make AI a team activity rather than a private browser tab, Claude Tag is the most credible thing a model provider has shipped. We build ChannelGate, which solves an overlapping problem, so treat this page as what it is: a comparison written by one of the two vendors. We have tried to make it one you can check. Every claim about Claude Tag below comes from Anthropic's own announcement, and where Claude Tag is the better product we say so.

How does an AI agent in a team channel work?

Both products implement the same pattern, so it is worth understanding the pattern before comparing the two. If you have only ever used AI through a private chat window, this is a different shape of thing.

The agent lives in a channel, not in your browser. You mention it the way you would mention a colleague, and everyone in that channel sees the request, the work and the result. That is the whole point: the work is visible, so a teammate can pick up where you left off, correct a wrong assumption, or just learn how you asked. AI stops being something each person does privately and becomes something the team does where it already talks.

It reads the conversation for context. The thread is the brief. Nobody re-pastes the background into a separate tool, because the background is already there.

It remembers, scoped to that channel. What the agent learns in a sales channel stays in the sales channel. This is what makes it useful on the second day rather than a novelty on the first, and it is why scoping matters: shared memory without boundaries is a leak waiting to happen.

It acts through connected accounts and tools. Reading a file, querying a CRM, running code, posting a result. This is the step that turns a conversation into work, and it is also the step that makes access control the real product question rather than a footnote.

It keeps working when you stop typing. Scheduled runs, background jobs and follow-ups mean a task can outlive the message that started it.

An administrator decides the boundaries. Who can use it, which tools and accounts it may reach, what needs approval, what it costs, and what record is kept.

How a request moves through an AI agent in a team channel A four-stage flow. A request starts in the shared channel, where anyone can mention the agent and the thread acts as the brief. It passes to channel-scoped memory, where what the agent learns stays inside that channel. It then reaches approved tools and accounts, covering connected apps, files and code. The result returns to the same thread together with a record of what ran. Beneath all four stages, a band shows that an administrator sets the boundary: who may use the agent, which tools and accounts it may reach, what needs approval, and what it costs. 01020304 The channelScoped memoryApproved toolsBack to the thread Anyone mentions theagent. The thread isthe brief. What it learns herestays here, and not inanother channel. Connected accounts,files and code, withingranted permissions. The result, plus arecord of what ranand who asked. An administrator sets the boundary Who may use the agent, which tools and accounts it may reach, what needs approval, what it costs, and what record is kept.
The shared-channel pattern. Claude Tag and ChannelGate both implement all four stages; they differ in where stages two to four physically run.

Claude Tag and ChannelGate both do all six. They differ in two places: where it runs, meaning on Anthropic's infrastructure or on a Linux environment you own, and how much of the boundary is yours to set. Almost every real decision between them traces back to one of those two.

What does Claude Tag actually do?

These are not marketing paraphrases. Each one is stated in Anthropic's own announcement.

  • One Claude per channel. "Within a given Slack channel, there's one Claude that interacts with everyone." Work is shared by default, and a teammate can continue a task someone else started.
  • Memory scoped to the channel. Everything Claude learns, including its memories, stays scoped to the channels it has been given. What it picks up in a sales channel does not leak into an engineering one.
  • Tasks it schedules for itself. It can "schedule tasks for itself, pursuing a project autonomously over hours or days" rather than only answering when spoken to.
  • An ambient mode. Optionally, it proactively posts updates, flags information, and follows up on threads or tasks that have gone quiet.
  • An audit log. Admins can view a log of everything @Claude has done, along with who requested each task.
  • Per-channel spend limits. Administrators can cap token spend both for the organization and for individual channels.

That is a serious governance surface, and anyone telling you Claude Tag is an ungoverned toy has not read the announcement. If those six things describe what you need, the honest recommendation is to use Claude Tag.

What is Claude Tag better at?

A comparison that never concedes anything is not worth reading, so here is where we think Claude Tag wins outright.

Nothing to run. You enable it on a Claude Team or Enterprise plan and it works. ChannelGate needs a Linux host with rootless Podman and Node.js 22.13 or later, plus credentials for whichever engine you use. The install itself is three commands and the repo documents it as a ten-minute job, but you still have a service to operate, update and back up. If no one on your team wants to own that, the comparison is already over.

It acts without being asked. Claude Tag's ambient mode proactively posts updates, flags information and follows up on threads that have gone quiet. ChannelGate has configured follow-ups and nudges, which is a narrower thing: you set up the behavior, rather than the agent noticing a stalled thread on its own.

Hard spend caps. Administrators can cap token spend per organization and per channel. ChannelGate reports usage and cost by user, channel, engine and model, and limits runs by license tier, but those records inform review rather than stopping spend at a ceiling. If a finance team wants a hard limit, Claude Tag has the better answer today.

It is Anthropic's own product. The model provider built the integration, and that removes a layer of things that can break. For a team standardized on Claude with no plans to change, that alignment is a genuine advantage rather than a consolation.

Where does Claude Tag stop?

Three limits matter, and none of them is a flaw. They are consequences of it being a provider-hosted, provider-native product.

It is Slack-only. There is no Microsoft Teams or Google Chat version, and Anthropic has not named a roadmap platform. For a Microsoft shop this is decisive on its own.

It is Claude-only. Reasonably so, since it is Anthropic's product. But if your engineers want Codex for one thread and Claude Code for another, or you have a procurement position that says no single model provider gets the whole company, that is not a preference Claude Tag can express.

It runs on Anthropic's infrastructure. Your channel's memory, files and history live there. For most teams that is fine. For a regulated one, or one that has already committed to a data-residency position, it is the whole conversation.

One more thing is worth being precise about, because a lot of write-ups get it wrong: Anthropic has not published a billing model for Claude Tag. The announcement says launch credits are provided to eligible organizations. Usage-based billing is a reasonable inference, but if you see it stated as fact, the source is guessing. It is in beta for Claude Team and Enterprise plans.

Side by side

Where each product runs, compared Two architectures side by side. On the left, Claude Tag: Slack connects to Anthropic's infrastructure, which holds the agent, the channel memory, the files and the history, and which in turn reaches the Claude model. On the right, ChannelGate: Slack, and Microsoft Teams or Google Chat in Beta, connect to a Linux environment you own, which runs one rootless Podman container per conversation and holds the memory, files and history, and which in turn reaches your chosen model provider, either Anthropic or OpenAI. A note beneath both columns states that in each case the prompt and context still reach an external model provider, and that what differs is where the workspace, memory and history live. Claude Tag ChannelGate Slack Where the team asks Slack, or Teams and Google Chat in Beta Where the team asks Anthropic's infrastructure A Linux environment you own The agent runs hereChannel memoryFiles and historyAudit log One rootless Podman container per conversationChannel memory, as Markdown you can read and editFiles and history, in your deploymentAudit log, plus approvals held in the thread Claude model Your model provider: Anthropic or OpenAI In both, the dashed step leaves your walls: prompt and context reach an external model provider either way. What differs is where the workspace, the memory and the history live. Where each product runs, compared The same two architectures, stacked. Claude Tag: Slack connects to Anthropic's infrastructure, which holds the agent, the channel memory, the files, the history and the audit log, and which in turn reaches the Claude model. ChannelGate: Slack, and Microsoft Teams or Google Chat in Beta, connect to a Linux environment you own, which runs one rootless Podman container per conversation and holds the memory as Markdown you can read and edit, the files, the history, the audit log and approvals held in the thread, and which in turn reaches your chosen model provider, either Anthropic or OpenAI. A note beneath states that in each case the prompt and context still reach an external model provider, and that what differs is where the workspace, memory and history live. Claude Tag Slack Where the team asks Anthropic's infrastructure The agent runs hereChannel memoryFiles and historyAudit log Claude model ChannelGate Slack, or Teams and Google Chat in Beta Where the team asks A Linux environment you own One rootless Podman containerper conversation Channel memory, as Markdownyou can read and edit Files, history, audit log, approvals Your model provider: Anthropic or OpenAI The dashed step leaves your walls either way. What differs is where the workspace and memory live.
Self-hosting moves the middle block, not the model call.
Claude TagChannelGate
Effort to startEnable it on an eligible plan, nothing to runA Linux host with rootless Podman and Node.js 22.13+, then a documented ten-minute install, then a service you operate
Acting unpromptedOptional ambient mode posts updates and follows up on quiet threadsConfigured follow-ups and nudges, set up rather than self-initiated
Chat platformsSlackSlack (GA); Microsoft Teams and Google Chat (Beta)
AI agentClaudeClaude Code or Codex, chosen per thread, plus OpenCode as an experimental read-only engine
Model choiceClaudeModel and reasoning effort selected per thread, channel or gateway default, from whatever your connected provider account offers
When an engine failsNot applicableFallback you configure: off, automatic, or ask the author in the thread. An explicit pin blocks the switch
Where it runsAnthropic's infrastructureYour own Linux cloud or server, rootless Podman
MemoryScoped per channelScoped per channel, stored in your deployment as Markdown you can read and edit
Shared workOne Claude per channel, anyone continuesOne agent per channel, anyone contributes, steers or stops a run
Scheduled workSchedules tasks for itselfSchedules, background jobs and follow-ups that report back to the channel
AuditLog of every action and who requested itLog of every action, plus approval requests held in the thread
CredentialsManaged by AnthropicPersonal vs. shared channel connections kept separate; write-only channel secrets
Spend controlsPer-organization and per-channel token spend capsRun limits by license tier, plus usage and cost records by user, channel, engine and model. These inform review; they are not a spending cap
AvailabilityBeta, Claude Team and EnterpriseSlack GA; free Registered license, Enterprise by agreement
BillingNot published; launch credits for eligible organizationsFree tiers published on the ChannelGate licensing page; Enterprise quoted

The four questions that decide it

Rather than a feature count, these are the questions we see actually settle the choice.

Does the work have to stay in your environment?

If the answer is yes, because you are in a regulated industry, or you have made a data-residency commitment, or a security team has to sign off on where conversation history lives, then a provider-hosted product cannot be made to comply and the comparison ends here. ChannelGate is deployed by you on your own approved Linux environment, one rootless Podman container per conversation, and we do not host your work. One caveat we would rather state than have you discover: self-hosting moves the workspace, the memory, the files and the history into your environment, but the prompt and context still go to whichever model provider you connected, and connected apps still reach their own services. The boundaries of that isolation, and what sits outside them, are written out on the ChannelGate security page rather than summarized into a slogan.

If the answer is no, this is a point in Claude Tag's favor, and a real one. Zero infrastructure to run is an advantage you should weigh as such.

Do you need more than one agent?

Claude Tag is Claude, and that is the end of the question. ChannelGate picks Claude Code or Codex per thread, and within the chosen engine you also set the model and the reasoning effort. The honest limit is that this is choice within Anthropic and OpenAI rather than a free hand over every model, since what you can select depends on the provider account you connected. There is a third engine, OpenCode, but it is admitted only under a read-only, network-off profile, so it reads and searches a workspace and does nothing else.

Teams that care about this usually care for one of two reasons: their engineers have per-task preferences, or their procurement does not want a single-provider dependency. If neither applies to you, this is not worth paying for.

Is everyone actually in Slack?

Claude Tag is Slack-only. ChannelGate is generally available on Slack, with Microsoft Teams and Google Chat in Beta, and Beta means exactly that: the transport is implemented and tested, and it wants tenant validation before you rely on it. If you are a Microsoft shop, that Beta is still the only one of the two options that can reach your users.

Whose account does the agent act as?

This is the difference that is hardest to see in a feature table and matters most in practice. In ChannelGate, a request made by a person uses that person's connections, while shared channel connections are a separate and deliberately chosen thing, and channel secrets are write-only through the management interface. Sensitive actions can be held for approval in the thread before they run, and who may approve what is something you configure rather than something the platform decides for you.

Claude Tag is not careless here, and its audit log is genuinely good. The difference is that ChannelGate makes the identity an agent acts under an explicit, configurable decision rather than a property of the platform.

So which should you pick?

Choose Claude Tag if your company is on Slack, standardized on Anthropic, comfortable with Anthropic hosting the work, and you want something running this afternoon with no infrastructure. It is the lower-effort option and it does the core job well.

Choose ChannelGate if at least one of these is true: the work must stay in your own environment; you need more than one agent provider; some of your teams are on Teams or Google Chat; or you need per-identity credential separation and in-thread approvals. You will run a Linux host to get it, and that cost is real. If you would rather not run the rollout yourself, Makeitfuture does the ChannelGate implementation as a 30-hour engagement starting at €5,000.

A third honest answer: run both. They are not mutually exclusive, and a company evaluating how it wants agents to work is better served by comparing two real deployments than by comparing two marketing pages, including this one.

Frequently asked questions

What is Claude Tag?

Claude Tag is Anthropic's Slack integration, launched on 23 June 2026, that puts one shared @Claude into a Slack channel. Everyone in the channel works with the same Claude, its memory is scoped to the channels it is added to, it can schedule its own tasks, and admins get an audit log and per-channel token spend limits. It is in beta for Claude Team and Enterprise plans.

How much does Claude Tag cost?

Anthropic has not published a billing model for Claude Tag. The launch announcement states that launch credits are provided to eligible organizations, and access requires a Claude Team or Enterprise plan. Usage-based billing is widely assumed but is not a published fact as of September 2026.

Is there a self-hosted alternative to Claude Tag?

Yes. ChannelGate is deployed on your own Linux cloud or server environment using rootless Podman, so channel memory, files, skills and conversation history stay in infrastructure you operate. There are also open-source projects in this space, including open-tag and open-managed-agents.

Does Claude Tag work with Microsoft Teams?

No. Claude Tag is Slack-only, and Anthropic has not announced support for other chat platforms. ChannelGate is generally available on Slack and offers Microsoft Teams and Google Chat in Beta.

Can Claude Tag use a different AI model?

No. Claude Tag runs Claude. ChannelGate lets each thread select Claude Code or Codex, and within the chosen engine you also select the model and the reasoning effort, as a gateway, channel or thread default. Codex uses its live model catalog and Claude its rolling aliases, so the available models follow the provider account you connected. OpenCode is supported as a third engine under a read-only, network-off profile. An explicit engine or model pin is respected and prevents automatic fallback.

Do I need to host anything to use Claude Tag?

No. Claude Tag runs on Anthropic's infrastructure and you enable it on a Claude Team or Enterprise plan, with nothing to install or operate. A self-hosted alternative such as ChannelGate needs a Linux host with rootless Podman and Node.js 22.13 or later, plus credentials for the engine you use. The install is documented as a ten-minute job, but you are then operating a service.

Is Claude Tag secure enough for a regulated company?

It depends entirely on whether your obligations allow the work to be processed on Anthropic's infrastructure. Claude Tag's own controls, meaning channel-scoped memory, a full admin audit log and spend limits, are substantial. What it cannot offer is a deployment inside your own boundary, which is usually the blocking requirement in regulated procurement rather than the feature set.