Repository navigation
Replies: 4 comments
|
On the "served app is arbitrary code and shouldn't reach T3's own credentials or session" consideration, two cheap checks cover most of it for anything T3 listens on at loopback, and they are the ones we settled on for the CDP-style front door in Mars (the browser I work on):
The mirror image on the browser side is a grant list: loopback is reachable only for exact origins granted at startup, so even a tab that is allowed to load |
|
My use case is reviewing a remote agent's frontend changes in T3's integrated browser: render the app locally, annotate DOM elements, and send feedback with element context and screenshots to the remote agent. This would give sharp text and responsive scrolling/input while retaining T3's annotation workflow. The reviewer browser only needs to run during review. The agent can keep using its own browser on the remote machine for autonomous testing; preserving the reviewer's tab after their computer disconnects isn't needed. It would also be useful to expose the same authenticated HTTP/SOCKS proxy to Playwright, so a separate full Chrome instance on the reviewer machine can test the remote app. Playwright already supports HTTP/SOCKS proxies and the Chrome channel. This would avoid setting up a separate tunnel for each dev-server/API port. Related work:
@maria-rcks, could you share why #8841 was closed, and whether you'd reconsider that approach for local human review with annotations and reuse by Playwright? A desktop-first version scoped to selected remote dev ports would already cover my use case. |
|
I've opened draft PR #27879 with an implementation, screenshots, and two-machine verification. It lets the desktop browser render remote localhost apps locally, with element annotations and screenshot feedback sent to the remote agent. A separate full Chrome/Playwright instance can reuse the same authenticated HTTP proxy with CONNECT, even after the integrated webview closes. The agent's existing remote browser remains unchanged. Verified on two real Macs through T3 Connect with the remote server bound only to loopback: frontend/API/HMR ports, cookies, redirects, uploads/downloads, and annotation/image delivery to the remote agent. Unapproved ports fail without contacting reviewer-local services. Maintainer approval of the direction and scope is still outstanding. The draft provides working code and evidence for that discussion. |
|
I assumed it would be automatic but I can't see the references I made to this discussion from #37309 to resolve this discussion |
Uh oh!
There was an error while loading. Please reload this page.
Problem
Agents often start HTTP servers on the environment: dev servers, documentation previews, report pages, small dashboards. When the environment runs on another machine, there's no good way to use that server from the device I'm working on.
My setup: the environment runs on an always-on machine at home. I connect to it from a laptop and a phone through T3 Connect. An agent runs
viteonlocalhost:5173and links it in the thread. I want to click the link and use the app.What I can do today:
The alternatives aren't always available:
So when I'm away from the environment's network, the result depends on what my current device can install, even though T3 itself works.
What would solve it
From whatever device I'm using, I can open an HTTP server running on the environment, and my device renders it directly, as ordinary HTTP rather than a video stream. The traffic goes over the connection T3 already has to the environment, uses T3's authentication, and nothing is exposed publicly.
This isn't a request to change agent tabs. The environment-owned tab from #15328, which I share with the agent, is the right tool for watching and taking over an agent's work. This is about a person using an app the agent started.
Considerations
Related
All reactions