Browzer

browzer is agent-browser with a trace. Same commands, same snapshot-and-click loop. Each session is saved so you can inspect it later or open a local HTML replay.

Live watching stays with agent-browser dashboard. Browzer is the bundle you keep after close.

Install

Needs Node 18+, agent-browser on PATH, and ffmpeg on PATH (for video).

npm i -g @agentuity/browzer

Confirm:

browzer help
agent-browser --version
ffmpeg -version

Install the Agent Skill so Claude Code, Cursor, Codex, Grok, and others treat browzer as the browser CLI when a recording is wanted:

browzer skills install

Quick start

export AGENT_BROWSER_SESSION=demo
export AGENT_BROWSER_IDLE_TIMEOUT_MS=0

browzer --session demo open https://example.com
browzer --session demo snapshot -i
browzer --session demo find text "Learn more" click
browzer --session demo close

browzer traces
browzer inspect demo
browzer replay demo

Every browzer invocation except dashboard, mcp, doctor, install, upgrade, and session metadata is the same argv you would have passed to agent-browser. Do not also run agent-browser on that --session.

How a trace is captured

On the first successful command for a session, Browzer starts:

• record start → session.mp4
• network har start --content text → network.har

While commands run, it appends:

• commands.jsonl — argv, duration, exit, stdout/stderr
• events.jsonl — stream command / result / url / tabs
• console.jsonl — console and page errors from the stream

On close it dumps console.json, errors.json, snapshot.txt, poster.jpg, stops recording and HAR, and writes summary.md.

Traces are written under ~/.browzer/traces/<id>/ unless you pass --output / -o or set BROWZER_OUTPUT. Session bookkeeping still uses ~/.browzer/active (override the home with BROWZER_HOME). Override the wrapped binary with AGENT_BROWSER_BIN.

browzer --output ./artifacts --session demo open https://example.com
browzer --output ./artifacts traces
browzer --output ./artifacts replay demo

Commands Browzer adds

Command	What it does
browzer traces	List traces
browzer traces --json	Same, JSON
browzer inspect [id]	Print summary.md (latest if omitted)
browzer replay [id|path]	Write replay.html next to the video and open it
browzer skills	List the bundled skill
browzer skills get	Print SKILL.md matching this CLI
browzer skills install	Copy the skill into ~/.agents, plus Claude / Cursor / Codex / Grok if those dirs exist
browzer skills install --project	Copy into this repo’s agent skill folders

Everything else is forwarded to agent-browser.

Replay

browzer replay                 # latest
browzer replay demo            # id, prefix, or session suffix
browzer replay ./20260917-185027-long-demo

The page is a static HTML file beside session.mp4. No app server. The URL bar follows the playhead. Steps seek the video. Console, network, errors, and snapshot sit behind Show details.

After a failure

Agents should read browzer inspect first (failed commands, HAR 4xx/5xx, console errors). Humans open browzer replay. Live CDP-style inspection during a run is agent-browser dashboard start at http://localhost:4848.

Skill

Canonical file: skills/browzer/SKILL.md. Same Agent Skills format used by Claude Code, Cursor, Codex, and Grok. It is a prefix swap: same commands as agent-browser, invoke browzer. Snapshot/click/fill details stay in the agent-browser core skill.