aindeev/agent-chrome-relay

For developers who want Claude Code, Codex, Cursor or Paseo to finish web tasks that need an existing login, this is a macOS-only relay worth a look — if the local build and agent-allowlisting model fits your setup.

The problem

An agent asked to invite a teammate, check an admin panel or complete a checkout often stalls on a login page or a browser step it cannot reach. You end up copying values into a chat, or re-running the task by hand. Chrome Relay changes that step: the agent opens the page in its own Chrome tab, signs in with an already-signed-in account or with credentials from a shared 1Password vault, and then clicks through to the done condition. The result is a finished task rather than a partly executed plan.

What's different

The relay runs as a local background service (127.0.0.1:9333, installed as a LaunchAgent) and talks to a Go extension over chrome.debugger. That means no remote-debugging port, so Chrome never asks the user to allow remote debugging; it does show the "Chrome Relay started debugging this browser" banner while agents work. Each agent gets its own session and sees only the tabs it opened, and commands that could raise or focus the browser are swallowed, so the agent cannot pull the window to the front. The optional do command goes further: instead of one model round trip per click, the agent states a goal once and the relay carries it out, using saved flows, the agent's data, or Jev/Haiku under Claude Code, then hands control back. The trade-off is architectural: every browser-driving agent reaches one extension process inside one Chrome profile, and setup signs and loads that machinery on your Mac rather than on a shared server.

Sign-in assistance goes beyond simple password fills. When a Google flow offers an account that is already signed in, the extension clicks the account and Continue/Allow itself; otherwise it fills username, password and 2-step codes from 1Password vaults you share, including codes split across one box per digit. A code or link the site emails is read from the user's mail and typed or opened. The README's setup page and the one-prompt path point to two different setup routes: paste a setup prompt into Claude Code, Codex or Cursor, or clone and run ./build by hand.

Who should try it

A good fit is a macOS developer already using one of the four allowlisted agent apps who has recurring browser chores behind a company login — admin panels, internal tools, a heavily authenticated SaaS app — and is comfortable installing a signed local service. The 1Password integration changes the fit for people who keep credentials in shared vaults and want the agent to see item names, not values. A poor fit is anyone on Linux, Windows or a mixed fleet, or someone who needs a headless server-side browser; the project is explicitly macOS plus Google Chrome. It is also a weak match for teams that want to run agents in a container without a local Chrome profile.

Requirements and limitations

The first constraint is platform: macOS, Google Chrome and Go 1.25+ for the by-hand build, with Xcode command-line tools needed only for the 1Password vault and Secure Enclave passkey features. The second is access control: connections must carry a secret from ~/.chrome-relay/token and come from one of the allowlisted process names, checked via lsof/ps; browser-Origin connections are refused even with the secret. That is a deliberate narrowing, not a general-purpose remote-control API. Signing in with 1Password is off by default and must be enabled; the background-tab, Google sign-in and audit-trail parts are on by default and can be removed with chrome-relay uninstall. Ongoing upkeep means re-running the build after changes, keeping the LaunchAgent healthy, and managing which agent apps are allowed. The README does not document Windows or Linux support, so do not assume it.

HotRepo's take

HotRepo first recorded this project on 2026-10-24, when it appeared on the NEW-REPOSITORY board (新项目榜, not the overall daily Trending board) at #17 with 281 total stars. That is a single dated snapshot of a 14-day-old repository, so trend persistence cannot be judged from it. Before adopting, check the setup instructions against your agent app and confirm the Chrome Remote debugging banner behavior is acceptable to you.