No house model
Open Tag is an AI teammate you @mention in a channel. Every job runs on the model you chose for that job — Claude, GPT, Gemini, or open weights on your own hardware. We sell no inference, so we have no reason to send the work anywhere in particular.
Your keys · Your provider or your hardware · No token markupProvider rate-limited or deprecating a version? The same job re-runs on the fallback you set. The Slack workflow does not change.
Claude Tag is good. It is also Anthropic’s distribution channel for Claude, and that sets what it can be. This is economics, not a roadmap gap.
The model is the product
Anthropic launched Claude Tag on Opus 4.8, and a Slack teammate built by a frontier lab exists to run that lab’s model. Sending your channel triage to a 7B open model is revenue they would be choosing to give up. We have none to lose.
The weights are the moat
Claude Tag work runs in ephemeral sandboxes on Anthropic’s infrastructure, and it is not offered for third-party deployments. A closed frontier lab does not ship its weights to your data center. Open weights already run there.
The tokens are the revenue
Channel work is consumption-based, billed to an organization-funded balance on a Team or Enterprise plan. Open Tag never sits between you and a token: you hold the key and pay your provider directly, and we take no cut of what you spend.
Claude Tag details from Anthropic’s own materials as of August 2026 — launch note and support article. Product details change; the incentives behind them move slower.
Not a preference toggle buried in settings. The model is a property of the job, and you set it.
Cheap model for triage and summaries, frontier model for the hard reasoning, and a standing default per channel. One Slack identity either way, so nobody has to learn a second workflow.
Standing jobs keep running against the model you validated, instead of changing behavior the week your vendor ships an upgrade.
When a provider is degraded, rate-limited, or deprecating the version you depend on, the job moves to your fallback rather than waiting for someone else’s status page.
Point the sensitive channels at open weights on your own hardware and those prompts never leave your environment. There is no hosted sandbox in the path.
Cost and speed depend on the model, the workload, and where you host it. We publish the mechanism, not a blanket claim that open weights always win.
These are ordinary Slack jobs. What makes them ours is the second line of each one: the model that ran it, chosen by you.
Open Tag acts in Slack under the access you gave it. Admins and the person who asked keep the gates; everyone else just types @Open Tag.
Open Tag reads a channel only after a human invites it in. There is no workspace-wide scrape. Remove it and access ends with the membership.
Sending, spending, merging, or deleting stops at an approval. The gate is in the runtime, not in the model’s mood.
Integrations follow the person who connected them. The model is the one you configured. We do not train on your private content.
Same path as tagging a person. About two minutes, then a real job.
It only reads channels a human invites it into. Remove it and access ends with the membership. No workspace-wide ingest.
Bring your own keys — Anthropic, OpenAI, Google, or open weights on Ollama, vLLM, or your inference box. Set a default for the channel and a fallback for when a provider is down.
Ask for work, not a tutorial. It plans, runs, waits if something is irreversible, and posts the artifact back in the thread.
Stuck? Email support@opentag.bot with your workspace name.
Both live in Slack, both get tagged in a thread, both do the work instead of explaining it. Skipping that: here is what differs, and every line traces back to who owns the model.
| The structural question | Claude Tag | Open Tag |
|---|---|---|
| Who picks the model? | Anthropic. It launched on Opus 4.8 | You, per job, per channel. Claude included |
| Where does inference run? | Ephemeral sandboxes on Anthropic infrastructure; not offered for third-party deployments | Your provider, or your own hardware |
| Who bills the tokens? | Anthropic. Channel work is consumption-based against an org-funded balance | Your provider bills you. We take no cut |
| Can sensitive channels use it? | If you accept processing on Anthropic infrastructure | Yes, with local weights — prompts stay on your network |
| What happens when a model changes? | You move with the vendor’s versions and plans | Pin, fail over, or migrate. Slack workflow unchanged |
| What does it cost to leave? | The teammate goes away with the Claude contract | Swap the endpoint. MIT-licensed, self-hostable |
Claude Tag details from Anthropic’s public materials as of August 2026 — launch note and support article. Product details can change; if any row here goes stale, tell us and we will fix it.
We would rather you know this now than three weeks into a pilot. Pick Claude Tag if:
Choose us when the model should be a decision you keep making — per job, per channel, and after the contract is signed.
An AI teammate you tag into Slack. It has the context you give it in that channel, talks to the model you configured, and posts finished work: answers, files, and changes in connected tools.
The extra step is holding your own key, and it is the whole point. Anthropic decides what runs Claude Tag; you decide what runs Open Tag, including Claude. If you never want to make that decision, their product is genuinely simpler and we say so above.
Trust us with less. We are not in the inference path pricing your tokens, and we are MIT-licensed, so if we disappear the workflow is still yours to run. The model provider you already trust keeps doing the heavy lifting.
No. Most teams start with a hosted API key and never serve a model themselves. Local weights are there for the channels that cannot send data out — the option matters even if you never use it.
A chatbot tells you how. Open Tag does the job and leaves an artifact in the thread. You can still stop it, and anything irreversible waits for a person.
No. It reads a channel only after you invite it in. Remove it and access ends with the membership.
Yes. It can be wrong. It shows its plan so you can stop it, cites sources so you can check it, and asks before anything leaves your company that you cannot undo.
Content needed to run a job is processed so the teammate can finish it. We do not sell it, and we do not use your private content to train our own models. Your model provider — or your own box, if you run open weights locally — sees what you send it. Details are on Privacy and Security.
Invite it to a channel, point it at your models, and give it a job. Claude is welcome here — as an option you can change your mind about.