Skip to main content
A browser automation API should separate request handling from browser execution. Deploy the control API as a Springwinter Web server and deploy a bounded browser consumer as a Worker. Updated October 9, 2026. Return a job ID immediately; never keep an internet-facing request open for an unpredictable browser session.

​
API design

A small asynchronous API can expose:
The create endpoint validates the request, reserves quota, writes a job record, and publishes a queue message. The Worker creates the screenshot or PDF, uploads it to S3, and marks the job complete.

​
Prevent server-side request forgery

A user-supplied URL can target private infrastructure. Validate the initial hostname and every redirect. Resolve DNS, reject loopback, link-local, private, multicast, and metadata ranges, then protect against DNS rebinding by controlling outbound networking rather than trusting one lookup. Block cloud metadata endpoints, private VPC ranges, internal service domains, file: URLs, and unsupported protocols. An application-level URL check alone is not a complete egress boundary.

​
Bound every resource

Set limits for:
  • page and total job duration,
  • redirects and subresources,
  • response body and download size,
  • viewport and PDF dimensions,
  • concurrent pages per Worker,
  • jobs per tenant and per minute,
  • output retention,
  • browser restarts and retries.
Use an operation allowlist. Do not expose arbitrary Chrome DevTools Protocol commands to untrusted customers.

​
Deploy the components

Create a Web server for the API and a Worker for the browser queue. Store durable job state in a database and outputs in private S3. Use signed download URLs with short expirations. Monitor queue age, browser crashes, navigation timeouts, memory, output size, and completion rate. A healthy HTTP API does not prove that browser Workers are making progress.

​
Security limitation

The official Puppeteer Docker image expects SYS_ADMIN for its documented sandbox mode, but Fargate restricts that capability. Disabling Chrome’s sandbox is not sufficient isolation for arbitrary hostile pages. Treat a public browsing API as an agent-sandbox problem and choose a runtime that can enforce the required browser sandbox and network policy.

​
Frequently asked questions

It can for tightly controlled internal pages, but asynchronous jobs handle slow sites, retries, quotas, browser crashes, and client disconnects more safely.
Store them in private S3 and return short-lived signed URLs. The container filesystem is ephemeral and disappears when a task is replaced.
No. Protect every redirect and resolved address, block metadata and private ranges, account for DNS rebinding, and enforce network-level egress restrictions.

​
Sources and further reading

Last modified on October 8, 2026