The interface is settled.
The lock-in is not.
Something just got decided in how we work with AI agents. You don't open a separate chat app, paste in your code and your tickets and your context, and talk to an assistant in a room far from where the work lives. You mention an agent right there — in the issue, in the pull request, in the thread — and the result comes back in place. Claude Tag made that interface obvious, and it isn't going back.
We agree with the interface. We disagree with the cage it arrived in.
An agent that lives this deep in your workflow — reading your issues, touching your branches, answering in your threads — is not a casual tool. It sees what you see, and it acts where you act. Right now that same agent is quietly welded to one vendor's model, one vendor's cloud, and one vendor's idea of what you're allowed to inspect, run, or keep. The mention feels open. Everything behind it is not.
That trade looks fine on day one. It looks very different on day three hundred. The deeper this pattern spreads across your org — every repo, every channel, every team — the more of your work runs through someone else's runtime, and the harder it becomes to ever leave. Convenience compounds into dependency, and dependency compounds into lock-in. By the time it matters, the switching cost is the product.
So OpenTag is the open version.
Not a different interface — the same one. You still mention an agent where the work happens. You still get the result back in the thread. What changes is everything underneath:
The mention is a protocol, not a product. Point it at Claude Code, at Codex, at a smoke-test echo runner, or at an agent you wrote yourself — they all speak the same claim-and-callback contract. The model is a choice you make, not one made for you.
Run the daemon locally with opentagd and the work happens in your own checkout, against context you scope yourself — nothing gets copied into someone else's workspace. Prefer hosted? Run hosted. The point is that it stays your decision, on the same contracts either way.
Bindings, leases, context packets, approvals, apply plans, audit trails — first-class objects, not fine print. A runner claims only the work it is explicitly bound to handle. Changes arrive as proposals you approve before they land. Every step is on the record.
Humans get the acknowledgement and the result; the noise stays in the audit log. An agent in your workflow should feel like a good teammate, not a notification firehose.
OpenTag is not another AI workspace, and it isn't trying to win the interface — that fight is over. It's the layer that keeps the interface honest: a thin dispatcher between a mention and a runner, MIT-licensed, adapter-extensible, and unaffiliated with any vendor whose model you happen to use.
The interface won. Make sure you still own what's behind it.
Bring your own model. Keep your own context. Own your own agents.
← Back to home