Rysh vs Claude Tag: whose infra, whose governance, whose keys
Anthropic put Claude in your Slack. We love Claude — rysh runs on it. The real questions: where does the agent live, and who gates what it says?
Anthropic shipped Claude Tag — an org-level Claude agent living directly in your Slack workspace. First-party, hosted, zero ops. (It entered beta in June 2026; the older per-user "Claude in Slack" experience retires August 3.)
Let's get one thing out of the way: we love Claude. Rysh runs on it. Your rysh agents call the same Anthropic models, with your own API key. This is not a "our model is better" post — there is no our-model.
The comparison is about three other questions. Where does the agent live? Who governs what it posts? And what can it actually reach?
(All Claude Tag claims below are sourced from Anthropic's official docs as of July 2026 — verify current behavior before deciding.)
Question 1: who approves the message?
This is the big one.
Claude Tag posts and acts autonomously in channels. The official docs describe no human-approval gate before it posts — the admin docs even note it "may respond to a message that doesn't mention it when it judges a reply is warranted." The old private draft panel is the "Legacy" experience being retired. For many orgs, an assistant that speaks freely in company channels is fine. For plenty of others — regulated teams, customer-adjacent channels, week one of any AI rollout — it's the exact thing keeping the project stuck in review.
A rysh humanoid has a per-channel governance switch:
governance human— every reply is a draft. A person reads it in the terminal and confirms before it posts. Nothing reaches the channel until then.governance ai— autonomous, but every message still flows through a pane you can watch.
Flip it live, per channel. Start careful, earn trust, loosen where it's proven. Draft-and-confirm is live on email and website chat today.
Question 2: whose infra, whose keys?
Claude Tag is Anthropic-hosted only — Team/Enterprise plans, org credentials, no self-hosting, no third-party deployment. That's a feature if you want zero ops. It's a blocker if your security review asks "where does this run?" and the only answer is "not here."
A rysh humanoid is a process on your machine or server. Self-hostable (note: self-hostable, not open source — we don't blur that), bring your own Anthropic key. Your infra, your data path, your audit story.
Question 3: what can it reach?
Claude Tag has no access to your local machine; coding hands off to Claude Code on the web, which works against GitHub-connected repos.
A rysh humanoid sits next to your shell. Local tools, local repos, internal APIs — exactly as far as you allow, with dangerous operations gated behind approval. When someone asks the Slack agent "is the migration done?", a rysh humanoid can actually go look.
Question 4: how many agents do you want to maintain?
With rysh, the Slack presence isn't a separate product — it's a surface. The same markdown skill file that answers on Slack runs the same agent in your terminal, in your browser automations, and in your website chat. One definition, one audit trail — live in a pane, not only in an admin console. With Claude Tag, Slack is where the story ends.
Side by side
| Claude Tag | rysh humanoid | |
|---|---|---|
| Model | Claude (Anthropic-hosted) | Claude (your API key) |
| Posts autonomously | Yes — no pre-post approval gate in docs | Per-channel: autonomous or draft-and-confirm |
| Runs on | Anthropic's infra only | Your machine/server (self-hostable) |
| Local repo/tool access | No — hands off to Claude Code on the web | Yes — gated local tools, repos, APIs |
| Beyond Slack | Slack (Teams "coming soon") | Terminal, email, website chat, browser |
| Admin story | Org audit view, access bundles | Live panes + your own logs, your infra |
| Ops burden | Zero | You run a process |
Fair is fair: when Claude Tag is the right call
Genuinely:
- You want a hosted assistant in Slack with zero ops — nothing to run, nothing to patch.
- You're already on Claude Team/Enterprise and want first-party everything.
- The org-admin audit view and access bundles fit your compliance model.
- Anthropic's distribution means it'll be in a lot of workspaces fast, and polish will follow.
If "Claude, in Slack, managed by Anthropic" is the whole requirement — use Claude Tag. It's good.
When you want a governed employee instead
Pick rysh when the requirement sounds like this:
- "Nothing posts to a customer-adjacent channel without a human reading it first."
- "It has to run on our infra, with our keys."
- "The Slack agent should be able to check the actual deploy, not describe one."
- "We want one agent definition across Slack, email, web chat, and the terminal — not one product per channel."
That's not an assistant; that's an employee with a manager. It's what rysh humanoids are built to be.
We're early and building this with design partners. If your team wants governed agents on real channels — draft-and-confirm from day one, your infra — come shape it: become a design partner →
More in this series: Humanoids: agents with a Slack handle and an email address · governance ai|human: the draft-and-confirm switch