Claude Tag · Anthropic · Slack · 2026
The AI employee that lives in your team: works 24/7, has its own identity, and under the hood it runs Claude Code. With one catch — where its memory lives.
Not a marketing experiment
Not a demo: it's how they work internally. And it's not just them — Hebbia's CTO calls it their "first responder" for internal bugs, often leaving the fix ready. A GitLab director says it doesn't just do the work: it "challenges our thinking and provokes better questions". An employee that works 24/7, never forgets the company's context, and anyone in the channel can put to work just by naming it.
Karpathy called it the third move in how we use models. The change is subtle and enormous: the agent didn't arrive with a new tool — it changed location.
Web chat → you type, it replies
Desktop app → lives on your machine, helps you
IN YOUR TEAM → lives where everyone coordinates, runs on
its own, with its own identityMove 1
You type, it replies. The first way we used the model.
Move 2
It lives on your machine and helps you, one on one.
Move 3
It lives where everyone coordinates, runs on its own non-stop, with its own identity. The agent moved to where the work happens.
Where an agent lives IS what defines what it can do. We call that the harness — and today the harness is your organization. Watch Harness Engineering ›
What almost nobody is showing, and what changes a dev's mind. Anthropic itself says it plainly: Claude Tag is "an evolution of Claude Code".
"Claude Tag is the beginning of an evolution of Claude Code."
— Anthropic, official announcement
When you tag Claude in Slack and ask it to fix a bug or open a Pull Request, under the hood runs the same engine as your terminal: it reads the repo, runs the code, opens the PR, merges it. The model and capabilities didn't change — the harness did. The same Claude Code now lives where your repo, your tickets and your alerts are, with permissions an admin governs. That's why for a dev this isn't "another Slack bot": it's your coding agent, parked where the work is coordinated.
The most important idea technically. The agents we showed before (like Hermes or OpenClaw) acted as you: they used your keys, your account. In a team that breaks — whose credentials does it use? Claude Tag solves it at the root.
The agent lives per channel: access and identity follow the channel, not the person who calls it.
Its own accounts and keys, provisioned by an admin. Every action is logged with its credential — an audit trail.
It learns from the work, per channel, isolated. Private context doesn't leak into public.
Who accesses what, and the spend limit. The admin is in charge.
It's so literal you create its own access in each tool — its own email, its own user — just like when someone new joins: you don't hand over your login, you give them theirs. The official rule: "the credential you connect is Claude's account in that tool, not yours."
The limit by design
"Not because it's polite, but because it literally can't see it."
If you tag Claude in the legal channel to edit code, it can't: access follows the channel. It's the same layered-security idea from the Agent Skills video — you don't ask it nicely not to break anything, you remove the capability. Watch Agent Skills ›
Every request goes through the same cycle. The key: the sandbox is ephemeral and built per thread.
Start
A user tags @Claude in the channel, or a scheduled routine fires.
Sandbox
An isolated environment is built PER THREAD. Two threads in the same channel = separate sandboxes, no shared state.
Execution
It works with the access granted to the channel, updating a live checklist. It runs the same engine as Claude Code: reads docs, runs code, opens PRs.
Delivery
The result lands in the thread: a message, a file, a chart or a pull request.
Idle
The sandbox is released, the thread persists. A new reply rebuilds it from scratch. Only what was pushed out survives (a branch, a posted doc).
Not the list of twenty obvious things. The ones that actually change something, because the agent is proactive, has team memory and its own identity.
@Claude, we got three reports of the same bug this week. Investigate it and open the PR with the fix.
@Claude, review this PR and flag anything that could break in production.
@Claude, we need rate limiting on the API. Size the work before building it.
What truly didn't exist before: you don't have to call it, it raises its hand on its own. Official examples from Anthropic:
"That PR Diego opened Tuesday has gone two days without review, and the launch is tomorrow. Should we do something?"
"Watch this channel, classify what comes in by priority, and ping me only when something needs my decision."
Anthropic says usage spread to metrics, support and operations. Summarize the week's dashboard, triage tickets by urgency, or —a real Descript case— turn a whole season of event logistics into "a program that runs itself". The pattern is always the same: it lives where the work already happens, remembers the context, and acts with its own identity.
The honest part, what's worth knowing before diving in. Not to scare you — so you make the call with your eyes open.
It's not in your plan: it's billed by usage, with credits, to the organization. You need a Team or Enterprise plan (not Pro or Free). The first thing you configure is a spend limit — with an eager team, the bill runs up fast.
Today it runs only on Opus 4.8, replaces the old "Claude in Slack" integration with a 30-day migration window, and the behavior may change before GA.
The more you use it, the more it learns about your company — and that memory, the real asset, lives in Anthropic's house, not yours. The day you want to leave, the employee keeps everything it learned.
The employee is great, as long as you stay in the house where it lives. These are the four ways to solve identity and memory — and who owns it.
Service-account identity, channel-based access, memory in Anthropic.
Convenient, but locks you in
Personal OAuth: access and attribution follow the person.
Also in Anthropic
Local credentials: everything runs on your machine.
Yours, but individual
Your memory, in your cloud, multi-model, you govern it.
The next video
What if you could have all of this, but with a memory that's truly yours, without the lock-in? You can — and that's exactly what's coming in the next video.
Each connection is a dedicated account for the agent, not yours. It's not just Google Workspace — the list covers most of a team's toolbox.
Docs / knowledge
Google Drive · Notion · Confluence
Code
GitHub · GitLab
Data warehouse
BigQuery · Snowflake · Redshift
Monitoring
Sentry · Datadog · PagerDuty
Tickets
Linear · Asana · Jira
CRM / go-to-market
HubSpot · Salesforce · Gong
All data and quotes come from Anthropic's official documentation and announcements.
Anthropic · Jun 23 2026
The official announcement. 65% of the product team's code ships through Claude Tag, and usage spread well beyond engineering.
Anthropic · 2026
The official tagline: "@Claude reacts in real time, where the work is happening", and the Hebbia, GitLab and Descript quotes.
Anthropic · Docs · 2026
The 5-step loop, service-account identity, channel-based access and memory. The technical documentation.
Anthropic · 2026
Team and Enterprise plans, credit-based billing and the spend limit. What it really costs.
Claude Tag is the channel's thesis made product: the harness matters more than the model.
The agent didn't change capabilities, it changed houses. The harness is now your organization.
The same layered security: you don't ask it nicely not to break things, you remove its ability to see them.
The standardized connectivity. Claude Tag connects to your tools with a dedicated identity for each one.
The ambient mode that "raises its hand on its own" is an agent running in a loop, without you calling it.
Coordinated agents splitting the work — the pattern Claude Tag brings to the team's Slack.
Another surface of the same agentic pattern: the same engine living where the work happens.
The essentials about Claude Tag.
It's Claude inside Slack: you mention @Claude in a channel and it runs the task right there — reproduces bugs, opens pull requests, builds reports, investigates incidents or runs scheduled tasks. Anthropic launched it on June 23, 2026 in public beta, and under the hood it runs the same engine as Claude Code.
Anthropic describes it verbatim as "an evolution of Claude Code". Under the hood it runs the same engine: reads the repo, runs the code, opens and merges PRs. What changed isn't the model or the capabilities, it's where it lives: it left your terminal, where you used it alone, and entered the team's Slack, with identity and permissions an admin governs.
Memory accumulates per channel and lives in Anthropic's infrastructure. A public channel shares memory across the whole workspace; a private channel keeps it isolated. That memory is what makes it valuable over time — but also what locks you in: if you switch providers, the agent keeps everything it learned, in the previous house.
It requires a Team or Enterprise plan on claude.ai — it's not in Pro or Free. It's not part of your plan: it's billed by usage with credits, charged to the organization (DMs bill the personal seat). The first thing you configure in setup is a spend limit. There are launch credits, but they run out.
The agent has a service-account identity, not a personal one: it uses its own accounts and keys, provisioned by an admin, never yours. Access follows the channel, not the user, and credentials never enter the sandbox. Every action is logged with its credential (an audit trail). And if you tag it in a channel it doesn't have access to, it can't act — not because it's polite, but because it literally can't see it.
Claude Tag uses a service-account identity, access follows the channel, it has channel memory and bills the organization — it's for the team. Claude Cowork uses personal OAuth: access and attribution follow the person, and it bills your seat — it's personal. Same Claude underneath, different identity model.
No. It shines most in engineering — Hebbia's CTO calls it their "first responder" for internal bugs, often leaving the fix ready — but Anthropic says usage spread well beyond: metrics, support tickets, operations. Any team that works in Slack can use it.
Community
How to build an agent with its own identity and a memory that's truly yours, without vendor lock-in — step by step, with real repos to practice. Free community access; the full courses are in the Premium tier.
Join Agentic Engineers →YouTube channel
@NicolasNeiraGarcia
ADK · A2A · Claude Code · Automation · Infrastructure