Anthropic ships a Slack bot that writes 65% of its own product team's code

A Slack tag, sold as a teammate

Anthropic's announcement frames Claude Tag as more than a chatbot hook. You tag @Claude in a Slack channel; it breaks a request into steps, runs them with connected tools, and replies in-thread. The differentiator it leans on is multiplayer: one Claude per channel, visible to everyone, so a teammate can pick up a task where another left off. It builds memory from channel activity, can act proactively on stalled work, runs Opus 4.8, and is billed by token spend — a separate meter, not part of an existing Claude subscription. The proof point Anthropic puts up front: 65% of its own product team's code is created by an internal version of this tool.

That statistic became the thread's punchline rather than its proof.

"Today, 65% of our product team's code is created by our internal version of Claude Tag." "Yeah, that explains a lot." — a reader with 7,600+ karma reacting to the stat

The skepticism wasn't only snark. The recurring engineering objection was that 'multiplayer' collides with how permissions actually work.

The permission problem nobody in the demo answers

If Claude is one identity in a channel full of people with different access, whose permissions does it inherit? Several commenters circled this; one with a security background made it concrete.

"Claude needs to inherit permissions from somewhere, and those permissions will never align with the members of a slack channel. And finding the lowest common denominator of access probably results in a dumbed-down, useless experience." — a commenter on enterprise compliance

The most vivid version came from someone who builds exactly this kind of assistant for a living, at a calendar-scheduling company. In a private chat, an agent saying you can't meet at 11 because of a job interview is helpful. In a shared channel, it's a leak.

"imagine if one were to ask our service (or Claude) to find time with someone and it replied to the thread for everyone to see: 'The soonest I could find is 12pm tomorrow. Reggie is available at 11am, but Lightbody has a job interview so it won't work'. WHOOPS." — an engineer at a scheduling startup acquired by Dropbox

His conclusion: to be safe, you'd need multiple Claude identities, each scoped to what it's entitled to see — which reintroduces the very headcount-of-bots problem collaboration tools were meant to remove.

'Billed at API rates' — and the default is no spending cap

The second reservation was money. Because Claude Tag meters tokens separately, and because it can parse channel traffic to build memory, the cost surface is large and easy to miss.

"with Anthropic, every new feature has metered usage and 'unlimited spending' (aka no limit) enabled by default… if I activate something (Claude Code, Claude Tag) and don't actively go to the usage page to set a spending limit, there is no limit." — a deep-learning practitioner running a team org

Others noted the same shape from a different angle: this is roughly what you can wire up yourself with an open-weights model and a Slack app in an afternoon, except Anthropic does the setup and bills you at API rates. A founder building a competing 'AI teammate' product argued the genuinely hard part isn't the Slack plumbing at all — it's durable organizational memory plus access control that doesn't leak across a company.

Underneath the gripes sat a strategic read that several came to independently: the model labs want to own the interface, not just the model. Slack is the wedge; a few wondered aloud whether Anthropic eventually just builds its own. For now, the most grounded verdict in the thread was that this is a feature, not a product — and one whose unsolved question is printed in its own demo screenshots: who is Claude, and what is it allowed to know about each person in the room.