CLAUDE CODE
Claude Tag for Engineering Teams: A Practical Setup Guide
Claude Tag is Anthropic's persistent AI teammate inside Slack. You tag it, hand off a task, and it works asynchronously across a channel's shared context using the tools an admin connected. This guide is the part the launch coverage skipped: wiring it into an engineering team without leaking data or blowing the budget.
- What it is: Claude Tag, launched June 23, 2026, is a shared
@Claudeteammate in Slack that runs on Opus 4.8. It replaces the old Claude in Slack app under the same Slack app, so there is nothing to reinstall. - Setup is admin-led: an organization Owner provisions it at
claude.ai/admin-settings/claude-tagin four steps - create the identity, attach an access bundle, scope it to channels, then test. - Two decisions matter most: which tools each channel's access bundle includes, and whether ambient mode is on. Access follows the channel, not the user.
- Cost is governed, not guessed: billing is consumption-based, and an Owner sets a hard spend cap with a per-channel breakdown. Anthropic is giving a usage credit through September 1, 2026.
- What Is Claude Tag? (How It Differs from Claude in Slack)
- How to Set Up Claude Tag for an Engineering Team
- Connecting Claude Tag to Developer Tools and MCP Servers
- Ambient Mode: What It Reads and When to Enable It
- Controlling Claude Tag's Token Costs
- Claude Tag vs Copilot in Teams vs Terminal Claude Code
- Frequently Asked Questions
What Is Claude Tag? (How It Differs from Claude in Slack)
Claude Tag is a persistent AI teammate that lives in your Slack channels. You tag it with @Claude, describe a task in plain language, and it breaks the work into stages, runs through them using the tools it has access to, and posts the result back in the thread. Anthropic announced it on June 23, 2026 (Introducing Claude Tag), and it runs on Opus 4.8.
The old Claude in Slack app was a per-user assistant. You mentioned it, it answered you, and the context died with the thread. Claude Tag is a different shape. There is one shared @Claude identity per channel that every member can see and pick up from where the last person left off. It builds memory from channel history and connected data sources, so the team stops re-explaining the same context. With ambient mode on, it can act without being tagged. And it works asynchronously, scheduling tasks for itself and pursuing a project over hours or days.
The numbers Anthropic shared make the intent clear: 65% of its own product team's code is now created by their internal version of Claude Tag (Anthropic, 2026). The transition is friction-free on the install side. The same Slack app powers it, so existing setups keep working while you switch a workspace over (Claude Code in Slack docs).
The mental model that helped me: the old app was a tool you called. Claude Tag is a teammate you onboard. That single shift is why the setup work is about access scoping and cost limits, not installation.
How to Set Up Claude Tag for an Engineering Team
Setup is admin-led, not self-serve. An organization Owner provisions Claude Tag from claude.ai/admin-settings/claude-tag on a Claude Enterprise or Team plan, then scopes it channel by channel. The flow has four provisioning steps: create the identity, attach an access bundle of credentials and repos, choose which channels and workspaces it applies to, and test it. This is the documented order in Anthropic's admin settings.
The one detail that changes how you think about security: access follows the channel, not the user. Admins configure credentials per scope - a channel, a workspace, or the whole organization - and individual users cannot override what a channel is allowed to do. So the unit of permission is the channel, and your job is to keep each channel's bundle as small as the work demands.
For an engineering rollout, I would not start broad. Spin up a single private channel, give it a narrow bundle, and prove the loop before inviting Claude anywhere people actually work. Inviting it to a channel is a normal Slack command:
# 1. Create a private test channel for the rollout
/create #claude-eng-test
# 2. Invite the shared Claude identity into it
/invite @Claude
# Claude only responds in channels it has been invited to.
# Channel membership is itself an access-control layer.With the channel scoped and a tool bundle attached, hand Claude a small, verifiable task. The point of the test is not to be impressed; it is to confirm the access bundle resolves to the right repo and that the result comes back as a reviewable artifact, not a wall of text.
@Claude the checkout flow throws a 500 on empty carts in api/checkout.
Reproduce it, find the root cause, and open a draft PR with a fix and a test.
# Claude posts progress in-thread, then returns with a summary and
# action buttons: View Session and Create PR. Each coding task runs as a
# Claude Code session on the web, and can open one pull request.Two routing modes decide how mentions are handled, set from the Claude App Home in Slack. Code only sends every mention to a Claude Code session, which suits a dedicated engineering channel. Code + Chat lets Claude decide between a coding session and a chat answer, which suits a mixed channel where people also ask general questions (Claude Code in Slack docs). For an eng channel, I default to Code only and avoid the ambiguity.
Heads up on prompt injection: when you tag Claude, it reads the surrounding conversation and may follow instructions buried in other messages. Anthropic's own docs warn to only use it in trusted channels. Treat a channel's membership list as part of its threat model, the same way I argued in my post on hardening AI agents against prompt injection.
Connecting Claude Tag to Developer Tools and MCP Servers
Claude Tag is only as useful as the tools you connect to it. Admins attach credentials, repositories, and connections as part of a channel's access bundle, and the underlying integration surface is the Model Context Protocol. Claude Code's MCP ecosystem spans 300+ integrations, including GitHub, Sentry, Linear, and PostgreSQL, plus any custom internal server you build (Claude Code MCP docs).
The trick is to match each channel to the smallest set of tools that channel's work actually needs. An incident channel does not need write access to your planning board, and a planning channel has no business touching the production database. Because access follows the channel, you get to make that decision per channel instead of granting one giant bundle to everything. Here is the mapping I would start with:
# Not a config file Claude reads - a planning sketch for the access
# bundle an admin attaches to each channel at admin-settings/claude-tag.
#incidents: # on-call and outage response
tools: [sentry, pagerduty, github-read]
ambient_mode: on # proactive surfacing helps here
#pull-requests: # code review and small fixes
tools: [github-write] # can open one PR per session
ambient_mode: off
#sprint-planning: # roadmap and ticket grooming
tools: [linear, github-read]
ambient_mode: off
#ci-alerts: # build and pipeline failures
tools: [custom-ci-mcp, github-read]
ambient_mode: onThat last one, custom-ci-mcp, is where a lot of teams will get the most value: a server you write to expose your own pipeline. I have built MCP servers for exactly this kind of internal surface, like the Jenkins MCP server that lets an agent read build status and trigger jobs, and a Method CRM MCP server for business data. If you have a CI dashboard, an internal metrics API, or a deployment tool, wrapping it as an MCP server is what turns Claude Tag from a code-writer into something that understands your specific stack. My guide to the MCP code-execution pattern covers how to keep those servers efficient.
Ambient Mode: What It Reads and When to Enable It
Ambient mode is the feature that makes Claude Tag feel like a teammate rather than a command line. With it on, Claude proactively keeps you updated about whatever it thinks you might need to know, flagging information across the connected channel and tools without being tagged first (Anthropic, 2026). On an incident channel, that means it can spot a spike in Sentry and post about it before anyone asks.
The catch is the part TechCrunch flagged on launch day: to decide what to surface, ambient mode reads the channel's messages (TechCrunch, June 2026). That is a fair trade in an ops channel full of alerts and stack traces. It is a worse trade in a channel where people paste customer emails, credentials, or anything you would not want a persistent memory to retain. Memory is scoped to the channel, which limits the blast radius, but scoping is not the same as not reading.
My rule of thumb is simple. Turn ambient mode on for channels whose entire purpose is signal you want chased - incidents, CI failures, deploy logs. Keep it off everywhere people mix in sensitive content or think out loud. If you are unsure, leave it off and tag Claude explicitly. You lose the proactive nudge, but you keep full control over what it sees and when.
Controlling Claude Tag's Token Costs
Claude Tag is metered as model usage, on a consumption basis. That is the line item to watch, because an async teammate that schedules its own work over hours can quietly run up a bill in a way an interactive session never does. The good news is the controls are real: an Owner sets a hard cap on Claude's draw from the organization's usage balance, and there is a per-channel consumption breakdown at claude.ai/admin-settings/usage/claude-in-slack.
Anthropic is softening the start with an introductory usage credit running through September 1, 2026, so eligible orgs can find their real consumption pattern before it costs anything. Use that window deliberately. The mistake is to set a generous cap, forget about it, and discover the shape of your spend in a monthly invoice.
The discipline here is the same one I wrote about in tracking Claude Code costs: the bill is driven by volume, not by any single clever prompt. Start each channel with a low cap. Watch the per-channel breakdown for a week. Raise the cap only where the value is obvious, and leave it tight everywhere else. One channel doing real engineering work earns a bigger budget; a chatty general channel with ambient mode on does not.
One number to remember: channel and thread work bills the organization, while a direct message to Claude bills the individual user's seat and uses their own account. Keep team-funded work in channels so it lands on the org balance you are actually monitoring.
Claude Tag vs Copilot in Teams vs Terminal Claude Code
Claude Tag is part of a broader shift: team chat is becoming the control surface for AI agents in the enterprise. GitHub brought Copilot into Microsoft Teams, and OpenAI's Codex ships a Slack integration of its own. The question is not whether to put an agent in chat, but which one fits where your engineers already are and how it participates in your workflow.
The honest framing is that these are not all the same layer. Claude Tag is the async coordination layer - hand off a task in Slack, get a pull request or a summary back. Terminal Claude Code is the deep-work layer for real-time, interactive sessions on your own machine. Copilot in Teams centers on the GitHub platform with Actions, PR pipelines, and org policy. You do not pick one and retire the others; you place each where it earns its keep.
Claude Tag: shared async teammate in Slack, channel-scoped access and memory, opens pull requests via Claude Code, ambient monitoring. Best for hand-offs, incidents, and team-visible tasks.
Terminal Claude Code: interactive, full-context, real-time pair work on your machine. Best for focused building and complex refactors.
Copilot in Teams: agent in Microsoft Teams anchored to the GitHub platform. Best when your org already lives in Teams and GitHub Actions.
If your team already runs Claude Code, Claude Tag is the natural extension into shared workflows, and it pairs well with the agent patterns I covered in Claude managed agents and dynamic workflows. The async, multi-stage task model in Slack is the same one those posts explore, now wearing a teammate's name.
Start with one private channel, a minimal access bundle, ambient mode off, and a low spend cap. Hand Claude a small fix, review the pull request it opens, then widen access one channel at a time. The teams that get burned are the ones that grant everything on day one. Start from the official docs and provision from your admin settings.