Agent Quick Browse (AQB)

Agents finish multi-step browser tasks 1.6–4.3× faster than with Claude in Chrome.

Fast, cheap browser actions for agents. A decision model — TypeSafe's Jev, OpenAI's GPT-6 Luna Decisions, or Cloudflare's Clef — picks each action and element in your real, logged-in Chrome; a smarter agent plans the goals. No LLM in the loop. Runs in parallel background tabs, via CLI or MCP.

Recommended setup

• Model: Jev, text mode (the default with a gateway or TypeSafe key). Fastest and most reliable in the benchmarks below. With only an OpenAI key, GPT-6 Luna (text) is used, and is close behind.

• If Jev gets stuck on a long app flow, retry that sub-goal with --model luna or --model clef --text, or script it with --match commands.

• Screenshots (--vision, with --model luna or clef) only when meaning is in pixels — unlabeled icons, canvases, charts. It's slower and less reliable on normal pages.

• Driving agent: split tasks into sub-goals of a few explicit steps each, pass facts and files, check each result, and finish fiddly widgets (date pickers) yourself with browser_look + browser_click.

Related MCP server: agent-tab-bridge

Benchmarks

Agent tasks: Sonnet + AQB vs Sonnet + Claude in Chrome

Each task needs ~3 sub-goals. Same model (Sonnet) and prompt on both sides; average of 2 runs, all answered correctly.

With AQB the agent also made about a third as many tool calls (5.1 vs 14.8 per task).

Single goals: decision models alone

Time per goal, from start page to done (5 runs per decision-model cell, 1 run for Sonnet; pass rate shown when below 5/5). ✗ = failed every run.

Sonnet times exclude agent startup. Reproduce with bun bench/run.ts --levels 1,2,4,14,18,5 --reps 5.

App widgets: creator-studio patterns

Local copies of TikTok Studio / YouTube Studio widgets (bench/fixtures), checked against the exact end state. 5 runs per cell.

Long flows in a single goal are unreliable: split them into sub-goals, or script them with --match commands (bench/scripted.sh does the full draft in ~12s). Vision mode can't reach list items scrolled out of view. Reproduce with bun bench/run.ts --levels 22-29 --reps 5.

Scoping goals

• One call: one objective on one site with explicit steps — "search for X and open the result", "fill the form with these facts and submit", up to ~10 navigation steps, "open the story with the most points".

• Split up or do yourself: complex widgets like date pickers and Google Flights (unless you spell out every step), vague or judgment-heavy goals, anything needing writing.

• Pass every string to type (including dates) as facts, and files to upload as files. If a rich-text editor mangles typed text (hashtag/mention pickers), pass keystrokes (--keystrokes) to type it key by key. Check each result (status, reason, url, actions) before the next call.

Setup

Requires macOS, Google Chrome, Bun, and an API key for a decision model: a Vercel AI Gateway key (covers Jev and GPT-6 Luna), or a TypeSafe key (Jev) or OpenAI key (GPT-6 Luna) directly. Published on npm as agent-quick-browse. Paste this to your coding agent (Claude Code, Codex, Cursor, …):

Set up the Agent Quick Browse (AQB) MCP server (https://github.com/zachsents/agent-quick-browse) for me:

1. Run: bunx agent-quick-browse@latest setup
2. Tell me to open chrome://extensions, turn on Developer mode, click "Load unpacked" and pick
   ~/.aqb/extension. Wait until I confirm.
3. Check that one of these is exported in ~/.zprofile (login shells read that, not ~/.zshrc):
   AI_GATEWAY_API_KEY (Vercel AI Gateway, covers Jev + GPT-6 Luna), TYPESAFE_API_KEY (Jev), or
   OPENAI_API_KEY (GPT-6 Luna — then use --model luna / model "luna"). If none is, ask me to add one
   myself — never ask me to paste a key into chat.
4. Register the MCP server, launched through a login shell so it picks up that key. For Claude Code:
   claude mcp add -s user aqb -- zsh -lc "exec ~/.bun/bin/bun x agent-quick-browse@latest mcp"
   For other agents, add an MCP server with command `zsh` and
   args ["-lc", "exec ~/.bun/bin/bun x agent-quick-browse@latest mcp"].
5. Verify: bunx agent-quick-browse run "Open the Ask HN page" --url https://news.ycombinator.com
   It should print JSON with "status":"done" and a url ending in /ask.
6. Ask me whether to add this line to my global agent instructions (~/.claude/CLAUDE.md for
   Claude Code, ~/.codex/AGENTS.md or similar for other agents), and only add it if I say yes:
   "For routine web tasks (navigating, searching, filling forms), prefer the aqb MCP's
   browser_run over other browser tools; use those for visual work or when aqb is blocked."
7. Tell me to restart my agent session so the browser_* tools load.

Keys are read from the environment or from ~/.aqb/env (aqb setup --save-keys copies them there, so agents and non-login shells find them). AI_GATEWAY_API_KEY is used when set; otherwise TYPESAFE_API_KEY (Jev, default model) and OPENAI_API_KEY (--model luna) talk to the providers directly. Clef needs CLOUDFLARE_AI_ACCOUNT_ID + CLOUDFLARE_AI_API_TOKEN (a Workers AI token). After upgrading, re-run bunx agent-quick-browse@latest setup and reload the extension.

Usage

bunx agent-quick-browse run "search for Cloudflare and open its article" --url https://en.wikipedia.org --fact search=Cloudflare
bunx agent-quick-browse run "upload my resume" --url https://example.com/apply --file resume=/path/to/resume.pdf
bunx agent-quick-browse run "open my orders" --url https://amazon.com --model luna   # GPT-6 Luna Decisions (add --vision for screenshots)
bunx agent-quick-browse look --text     # what the model sees
bunx agent-quick-browse mcp       # MCP server: browser_run, browser_look, browser_click, browser_type, browser_key, browser_upload, browser_tabs, browser_attach, browser_release, …

Exact steps without a model, using the numbers from the latest look --text:

aqb click 12
aqb type 15 "Caption #chess #fyp" --keystrokes   # replaces the text and verifies it; add --submit to press Enter
aqb key Escape                                   # also Enter, Tab, ArrowDown, Cmd+A, Shift+Tab, …
aqb upload 16 ./video.mp4
aqb scroll down; aqb navigate <url>; aqb back; aqb text
aqb value 15                                     # a field's full current value
aqb click --match 'radio "No, it.s not made for kids'   # target by outline text, waiting up to --timeout 30s
aqb wait --match 'Uploaded'                      # wait until a line of the outline matches

For the short aqb command, install globally: bun add -g agent-quick-browse.

Each --session (or MCP session) gets its own background tab, in a tab group you name with --group (default "AQB"); omit --url to continue where it left off. Take over a tab you already have open with tabs + attach <tabId>. release ends a session: tabs it opened close (--keep leaves them open, ungrouped), and taken-over tabs are handed back to their original group. MCP servers release their sessions automatically when the agent disconnects.