Coshell x Notte: What a Coding Agent Can't Learn from Localhost
Coshell ships a localhost browser that answers one question: does it render? Every other question lives outside the sandbox. Here is why a coding agent needs a second browser, who gets to step into it, and what a Notte persona can finish that an inbox alone cannot.
Single-player coding agents made individual engineers faster and left their teams behind. One person runs the agent in their terminal, everyone else reads a summary afterwards. Coshell made the session itself multiplayer: several engineers prompting one agent, against one filesystem, one terminal, one running environment.
It ships with a localhost browser, so when the agent builds a web app the whole session watches the preview update. That browser answers one question: does it render?
The browser outside the sandbox
Every other question an agent needs answered is out there. Does this API return what its docs promise? What does the signup flow on staging look like right now? Has the latest release broken the interface I'm importing?
Those two browsers look alike and are not. The localhost one renders a process you own, deterministic and yours to restart. The second is an interface to systems you don't control, where pages move under you and failure is ambiguous by default. Did the check fail because the code is wrong, or because the page changed? That ambiguity is the expensive part, and it compounds. An unverified assumption from early in a run is still load-bearing at the end of it.
"It isn't a headless session in a log: it's a live frame in the shared session, so the whole team watches the agent browse, and when it hits a login or a judgment call, any teammate clicks in, handles it, and hands back."
Michael, Coshell
Where Coshell's team steps in
Plenty of cloud browsers will stream you a session, so watching is not the hard part. The harder question is who is allowed to intervene, and at what moment.
Agents reliably struggle in two places: an authentication wall, and a judgment call no model should make alone. Left alone at either, an agent stalls or guesses, and guessing at a login is the worse of the two. In Coshell the session is already shared, so whoever is in it can take the browser, clear the login or make the call, and hand it back.
What a Notte persona can finish
"Because Notte personas come with a working email address and phone number, the agent can do the one thing no founder can: sign up for the product as a complete stranger, verification email and 2FA text included, while the team watches its own first-run experience live."
Michael, Coshell
The mechanic is unremarkable, which is the point. The verification email lands in the persona's inbox, the agent reads it, confirms, and keeps going. No test accounts, no mocked flows, and no founder muscle memory quietly skipping the steps that confuse everyone else.
Where Notte actually fits
We are not the only cloud browser you could point at this. Browserbase and Kernel both run good managed sessions with live views, and Browser Use gives you the reasoning loop in your own codebase with an inbox attached. If your signup sends a link, any of them finishes it.
What we spent our time on is making the identity whole. A Notte persona holds an inbox and a phone number on the same object, so a signup that sends a text does not end the run. One identity, one lifecycle, one place it gets revoked. An inbox stapled on from a third party clears the link-based half of verification and stops at the half that carries 2FA. That is a narrow difference, and we would rather name it narrowly than dress it up as a category.
The other thing we will say plainly is speed. Browser Arena is our own benchmark, open source and reproducible, and it runs every major cloud browser through the same test on the same machines: create a session, connect over CDP, navigate, release. Notte connects a session in 353ms, the fastest of the seven providers on the board, at $0.05 an hour, and leads the composite of reliability, latency and cost. We publish the harness because a benchmark nobody can rerun is just a claim. When an agent opens a browser on almost every turn, session start stops being a detail and starts being the loop.
Coshell made agentic coding multiplayer. The second browser is how that team checks its work against the systems it doesn't control, with everyone watching. Writing the code and knowing it works are two different problems, and the second one is where the sandbox runs out.