playwright-browser

playwright-browser

A forkd parent built from mcr.microsoft.com/playwright — the official Microsoft Playwright image with Node.js + Playwright + Chromium (and Firefox/WebKit) + all dependency .so files preinstalled. The parent VM keeps a headless Chromium process alive through snapshot, so every child fork inherits the warmed browser via mmap CoW.

Status: working. End-to-end verified on a bare-metal i7-12700 dev box: snapshot of a warmed Chromium parent → fork 3 children in 56 ms wall-clock → sb.eval("return await page.title()") returns in 10–82 ms per child. Two gotchas baked into the build flow below:

• Parent VM needs ≥ 2 GiB memory (Chromium OOMs at the 512 MiB default); use --mem-size-mib 2048 on forkd snapshot.
• Per-child cgroup ceiling should be ≥ 2560 MiB to stay above Chromium peak RSS.

Why this recipe

Browser fan-out is the second-largest AI-agent workload shape after Python — Anthropic computer-use, OpenAI web browsing, every coding agent that uses Playwright/Puppeteer for in-browser interactions.

Cold-start of a fresh Chromium-in-container is 2–3 seconds:

Step	Cold-start cost
Container start	~300 ms
node boot + Playwright lib load	~400 ms
chromium.launch() (CDP, renderer process)	~1.2 s
First newPage() + goto(about:blank)	~100 ms
Total per fresh browser	~2 s

With forkd, the parent VM does this work once at snapshot time; every child fork inherits the post-launch state in ~10 ms. 100– 300× faster per browser instance.

This is the workload shape Anthropic's computer-use and OpenAI's browser tool are on — many short-lived, parallel browser sessions, each needing a fresh isolated context.

What you get

• mcr.microsoft.com/playwright:v1.50.0-jammy base
• Pre-launched headless Chromium with one about:blank page, resident in parent memory at snapshot time
• forkd-init.sh + forkd-agent (Node bridge — landing) running as PID 1; agent exposes eval / page.* / browser.* over TCP

Total rootfs: ~2.5 GB, memory image after warm-up: ~1.5 GiB.

Use it

sudo bash recipes/playwright-browser/build.sh
sudo bash scripts/host-tap.sh
sudo -E forkd snapshot --tag pwb \
    --kernel ./vmlinux-6.1.141 \
    --rootfs recipes/playwright-browser/parent.ext4 \
    --tap forkd-tap0 \
    --boot-wait-secs 25      # Chromium renderer init takes longer than Python import
    --mem-size-mib 2048      # Chromium OOMs on the 512 MiB default

# Fork 50 browser sessions, all share the warmed Chromium
sudo bash scripts/netns-setup.sh 50
sudo -E forkd fork --tag pwb -n 50 --per-child-netns --memory-limit-mib 2560

# Drive one of them via the warmed Chromium
sudo -E forkd eval --child forkd-child-7 -- \
    "await page.goto('https://example.com'); return await page.title()"
# → "Example Domain"   (typical: 10–80 ms)

Python SDK

from forkd import Sandbox

with Sandbox(tag="pwb") as sb:
    # Browser is already warm — no Chromium launch cost. The agent
    # routes `eval` to the warmed Node + Playwright in PID 1's child;
    # `page`, `context`, `browser` are in scope.
    title = sb.eval(
        "await page.goto('https://example.com'); return await page.title()"
    )
    print(title)  # → "Example Domain"

When to pick this

• You're building an AI agent that drives a browser (computer- use, web-research agent, scraping agent, end-to-end UI test generator).
• You run Playwright test suites at parallel scale and pay multi-second-per-browser cold start.
• You want per-task browser isolation without the Docker cold-start tax.

When NOT to pick this

• You only need Python without a browser → use python-numpy/ (1/2 the size).
• You want the full IDE + VSCode + browser kitchen sink → use agent-workbench/.
• You need to drive a real GPU-accelerated browser (forkd children share the parent's headless config; switching to --enable-gpu per-child needs a different warmup pattern).

Benchmarks

To be filled in once the recipe's end-to-end Firecracker run is verified. Target shape: 50 concurrent fresh Chromium pages reachable in <500 ms wall-clock, vs ~100 s cold-boot Playwright-in-Docker.