Anthropic Just Put Claude in Your Slack — And It Wants to Be a Coworker
Tuesday morning, 9:47 AM. My Slack pinged with an Anthropic update. It wasn’t a new model. It was Claude Tag.
The pitch: in any Slack channel, you can now @Claude like you’d @ a teammate. You give it a task. It does the task. You can also direct-message it, and it works with the personal tools and connectors you’ve already set up.
First reaction: this is the first time Anthropic has said out loud that they want Claude to be a coworker, not an assistant. That sounds small. It isn’t.
What you can do with @Claude
The feature set, briefly. You can set Claude a task and walk away — it’ll work in the background. You can also schedule tasks that run autonomously over hours or days. Multiple Claudes can run in parallel, each scoped to a different channel.
Admins control the scope. Each channel has its own memory, its own data access, its own tools. A Claude set up for sales doesn’t see engineering data. An engineer, Claude, doesn’t get sales tools. Memories stay isolated.
Anthropic says they use it internally and “now spend much more of our time delegating tasks to many Claudes in parallel.” That’s a real admission from a frontier lab. They’re not just selling the tool — they built it for themselves first.
The launch includes a video walkthrough, a launch credit, and a connection to Claude in Slack. There are also docs on how to set up admin provisioning. This is a real product, not a teaser.
Why Anthropic is doing this now
Two reasons that don’t get talked about enough.
First, every AI lab is racing to be the layer where work happens. OpenAI has ChatGPT, Anthropic has Claude.ai, and Google has Gemini. The model matters less every quarter. The surface where the model meets your day matters more. Slack is where a lot of work already happens. Anthropic wants to be there.
Second, the underlying tech is finally good enough. The intro page says Claude Tag uses the same Claude models as the rest of the product. What’s new is the agentic plumbing — long-running tasks, parallel execution, scoped memory. None of that worked reliably in 2024. It mostly works now.
The HN announcement post hit 268 points in five days. For a product launch that doesn’t include a new model name, that’s unusually high. ZDNet, TheNewStack, and PromptQL all wrote teardowns. The developer audience is paying attention.
What’s still missing
A few things, from the public docs and what people are saying in the comments.
There’s no mention of pricing beyond the launch credit. That’s likely to change in a few months. There’s no public benchmark on how reliable the long-running tasks are — that’s the part of agentic systems that breaks most often in my experience. The admin setup looks involved, which means smaller teams will probably skip it.
The bigger question is whether teams will actually want this. Slack already has too many bots that ping you about things you didn’t ask for. A coworker who never sleeps is either the future of work or a way to make everyone anxious about their job. Probably both.
I’ll be testing it this week. If I find something surprising, I’ll write it up. If I don’t, I won’t.
Last updated: 2026-06-28
