GitHub - ryanbrunner/gloss

Review a running web app in a browser with a feedback bar across the top, and hand what you wrote to Claude.

gloss open <url> opens the page in a Chromium window with the Gloss bar over it. The reviewer writes comments in the bar and presses Submit to send them to Claude as a round. Claude applies them and the window reloads with its summary of what changed. This repeats until the reviewer presses Approve.

Gloss never fixes anything itself. It is the channel between the reviewer and whatever runs Claude (Claude Code, or Reeve). The agent drives the session with four short commands: gloss wait for the reviewer's verdict, gloss working and gloss ready either side of its changes, and gloss close. The Claude Code plugin in this repo runs that loop for you.

Install

With Homebrew:

brew install ryanbrunner/tap/gloss
gloss install-chromium            # once: about 150 MB

gloss install-chromium downloads the Chromium that Gloss's own Playwright drives, so the browser always matches it. On Linux it also needs system libraries that the download does not include; gloss install-chromium says so rather than installing them unasked. Run gloss install-chromium --with-deps to get those too (it asks for sudo).

From a checkout

Node 22.12 or later.

npm install
node bin/gloss.js install-chromium   # once: about 150 MB
npm link                             # optional: puts `gloss` on your PATH

Without npm link, run it as node bin/gloss.js … or npm run gloss -- ….

The Claude Code plugin

The repo is a Claude Code plugin marketplace. With gloss on your PATH:

/plugin marketplace add ryanbrunner/gloss
/plugin install gloss@gloss

From a checkout, /plugin marketplace add /path/to/gloss works too.

Then /gloss http://localhost:3000 opens the page and loops: it waits for your round, marks itself working, applies the comments, says it is ready with a summary, and waits again, until you approve.

Commands

gloss open <url> [--name N]    show <url> in the Gloss window, starting a session if there is none
gloss status [--name N] [--json]   exit 0 and say where the session is and its phase, or exit 1 if there is none
gloss close [--name N]         end the session and close its window
gloss install-chromium [--with-deps]   download the Chromium Gloss drives, once
gloss --version                print Gloss's version

gloss wait [--name N]          block until the reviewer submits or approves; print the verdict as JSON
gloss working [message]        the bar says Claude is working, with the message; the reviewer cannot submit
gloss ready [summary]          reload the window, show the summary, and hand the next round to the reviewer

gloss wait prints one JSON document on stdout and nothing else, described in docs/verdict.md. Only exit 0 with "approved": true is approval. Exit 1 with nothing on stdout means no verdict: no session, or the window was closed or the session ended while it waited. Asked again before working or ready, it prints the same round, so after a round the next command is gloss working, not gloss wait.

A round goes: the reviewer submits (submitted), the agent runs gloss working (working), then gloss ready (back to reviewing, one round on). Approve ends it (approved).

gloss open returns as soon as the window is up (within 15 seconds) and prints the session's address. Run it again from the same directory and it moves the same window to the new URL. Closing the window ends the session too.

There is one session per working directory. --name gives a directory more than one.

Where things live

• ~/.gloss/sessions/<id>.json: the running session's pid, port and token. The id is a hash of the directory and --name. It is outside the repo, so a session never dirties a worktree.
• ~/.gloss/logs/<id>.log: the session process's output.

Set GLOSS_HOME to move both.

The session API

The session serves a small HTTP API on 127.0.0.1 only. Every request needs the token from the state file:

state=~/.gloss/sessions/<id>.json
curl -H "Authorization: Bearer $(jq -r .token $state)" \
  "http://127.0.0.1:$(jq -r .port $state)/api/state"

round is the round being written now. phase is reviewing, submitted, working or approved. A comment's sentIn is the round it went out in, or null while it is unsent. A pinned comment's pin has the element's selector, tag, text, box and the viewport, the path of its screenshot, and quote, when it was made on a selection.

The bar

• Type a comment and press Enter or Add. Shift+Enter adds a new line.
• Interact and Select, beside the round, are the tools. In Interact the page works as usual. In Select, the page's links and buttons do nothing: the element under the pointer is outlined, with its tag and size, and a click opens a comment box beside it, headed with the element's tag and text. Enter or Add pins the comment to the element; the session keeps a selector for it and a screenshot of it, taken with the outline and markers hidden. Select stays on after Add or Cancel, for the next element, until you press Escape (which closes an open box first) or Interact. On a phone the tools are one crosshair button that turns Select on and off. They are disabled once the review is approved.
• In Interact, selecting text on the page offers a Comment on selection button beside it; the comment it starts is pinned the same way, with the selected words kept alongside it as a quote.
• Each pinned comment gets a numbered marker on its element's top right corner, which follows it as the page scrolls; hover it to see the comment and outline the element. The list shows the same number with the element's tag and text. A marker for an unsent comment stays where the element was if its selector stops finding it. Once a round is sent, its markers are dimmed, and shown only where the selector still finds a visible element.
• Comments (n) counts the comments not yet sent. The list shows those first, each with a delete button, then every earlier round under "Sent in round N", dimmed and read-only, so you can check what was asked.
• Submit sends the unsent comments as a round. It is disabled when there is nothing new, and while the round is with Claude.
• Approve ends the review. With unsent comments, it asks first, in the bar: "Approve and discard N unsent comments?" Discard & approve deletes them for good (they never reach Claude, and there is no undo); Cancel leaves everything as it was.
• The status beside the buttons says where the round is: "Sent round N, waiting for Claude", "Claude is working: …", Claude's summary once it is ready, or "Approved". On a phone it sits in a line under the bar.
• Comments belong to the session, not the page, so they survive a reload and every tab shows the same list.
• The bar lives in a shadow root on one <gloss-bar> element on <html>. Page CSS cannot reach into it and its CSS cannot leak out. The changes to the page's own styles are html { margin-top: 44px }, which pushes the page down below the bar, and 44px added to the page's own scroll-padding-top, so an anchor jump lands below the bar and any sticky header the page allows for. Fixed and sticky elements placed from the top of the viewport, such as a header, are moved down by the same 44px. And since 100vh still measures the whole window, each vh length in the page's stylesheets is shortened to match: 100vh becomes calc(100vh - 44px), so a full-height layout ends at the bottom of the window rather than 44px past it.
• The page under review is the code Claude is editing, so it must not be able to approve. The bar's script runs in an isolated world: it shares the page's DOM but none of its JavaScript, and only that world can reach the session. Nothing the page patches (JSON, Map, Promise, a prototype) is on the path a call takes. Submit, Approve and a comment's page link act only on trusted clicks, so the page clicking them through the shadow root does nothing, and the session refuses changes from frames and from pages that are not http or https. scripts/spikes/loop-check.ts tries each of these from the page.

Known gaps

• A vh length the bar cannot rewrite still overflows by 44px: one in an inline style attribute (including a --vh the page sets from innerHeight), a cross-origin stylesheet, the page's own adopted or shadow-root sheets, or a rule inserted into a sheet after it loaded, as CSS-in-JS libraries do in production. vmin, vmax and min-height media queries still measure the whole window.
• A fixed or sticky element inside a web component's shadow root is not moved, and sits under the bar.
• An element inside a web component's shadow root is pinned as the component itself.
• A pin's selector is the path to the element when it was picked (#summary > p:nth-of-type(3)). After the agent's changes it may find a different element, and a sent pin's dimmed marker then sits on that one.
• The window is Chrome for Testing, not your own browser: it has no profile, logins or extensions.
• document.execCommand edits whatever has focus, including the comment box, and in Chromium fires only a trusted input event with no beforeinput first, unlike a real edit. createValueGuard in src/bar/dom.ts marks a trusted edit as the box's own on its beforeinput, after every page capture listener further up the tree has had its turn, and bar.ts stops beforeinput and input from bubbling past the shadow root, so a page bubble listener cannot run in the gap before the matching input either. What checks that input is on window, in the capture phase and registered before the page's own scripts run, so it is still the first to see one fired from inside a page's own capture listener's execCommand, nested in the real edit's dispatch. It puts the box's value back whenever an input arrives on a box not expecting one, so the page cannot rewrite the comment the reviewer is about to send. A page task already queued when a beforeinput gets no input at all - backspace in an empty box - could still land in the small gap before the fallback clears that box's flag. It cannot stop a page that writes value straight through the shadow root, which is still open; that wants the root closed.

Why Playwright, not a proxy or an iframe

To pin comments to elements, the bar needs to reach the page's DOM. A cross-origin iframe can't do that, and X-Frame-Options or a CSP can refuse framing altogether. That leaves two choices: a proxy that injects the bar into the HTML it serves, or a browser that injects it for us. Gloss drives Chromium with Playwright, and gives every page the bar over DevTools: a script run on each new document, in an isolated world, and a binding only that world can call:

• The page is untouched. It loads from its real origin. A proxy would have to decompress and rewrite HTML and absolute URLs, relay the dev server's HMR websocket, strip CSP and frame headers, and keep cookies and redirects from escaping to the real origin.
• A strict CSP doesn't stop it. The script runs whatever the page's CSP says. The bar's styles are constructed stylesheets rather than <style> tags, and it talks to the session through the binding rather than fetch, so neither style-src nor connect-src gets in the way.
• It follows you. New tabs, client-side navigation and full reloads all get the bar again.
• HTTPS dev hosts load, with ignoreHTTPSErrors, as in Reeve's screenshots.

The cost is the Chromium download and a window that isn't your everyday browser. If Chromium is missing, gloss open exits 1 and says to run gloss install-chromium.

Development

npm test              # unit tests
npm run typecheck
npm run dev -- -port 4400
npm run schema        # rewrite schema/verdict.v1.json from src/verdict.ts
npx tsx scripts/spikes/open-check.ts   # end to end, headless; needs Chromium
npx tsx scripts/spikes/loop-check.ts   # the review loop: rounds, working, ready, approve, no-verdict cases
npx tsx scripts/spikes/nav-check.ts    # links, forms, client nav, X-Frame-Options: DENY

CI runs all three spikes on every push and pull request, in .github/workflows/ci.yml. It also installs the packed tarball the way Homebrew does and runs it on node 22 and 26, and lints the formula. docs/releasing.md covers releases.

npm run dev serves a fixture storefront to point gloss open at. It also reads --port and PORT. Query flags make each state of the bar reachable by URL, using the same bar with its comments kept in the page. &pins=N pins the first N comments, sent ones first, and &pick=N picks the Nth of the same elements; both take numbers, since a selector's # would end the query:

Two environment variables exist for the spike: GLOSS_HEADLESS=1 runs the session's Chromium headless, and GLOSS_CDP_PORT opens a DevTools port on it so the spike can look inside the window.