Limits and retention - Webcompute

Limits keep browser jobs predictable. Some limits are source-defined and public. Others can vary by account or deployment.

Public limits to design around

Area	Current guidance
Browser duration	Pass `maxDuration` when creating browsers that should expire sooner than the default.
Browser-code execution timeout	Keep snippets bounded. SDK and MCP surfaces cap execution timeout.
Staged files	The SDK rejects staged file uploads above its file size limit.
Downloads	Use `maxBytes` when saving or reading downloads.
MCP tool calls	`run_web_agent` accepts `maxTurns`, `maxToolCalls`, and `toolTimeoutMs` on the agent surface.
Recordings and artifacts	Store IDs you need, and close browsers when the workflow is done.

Use the limits reference for exact current values that are documented publicly.

Account-specific limits

Concurrency, rate limits, retention windows, and cost controls may depend on your account or deployment. Your application should handle 429, 413, timeout, and resource-limit errors explicitly.

Cost controls

Public source does not define a universal billing-control API in these docs. Treat cost as an application concern:

Set browser maxDuration for job-scoped browsers.
Set agent maxTurns, maxToolCalls, timeoutMs, and toolTimeoutMs.
Cap downloads, uploads, observations, and returned data.
Close browsers when a job is complete.
Log usage metadata when web.agent() returns it.

Cleanup checklist

Close browsers created for one job.
Store only the artifact IDs and summaries your app needs.
Avoid logging signed Debug UI or CDP URLs.
Put file size caps on download reads and saves.
Use policy to constrain uploads and downloads.

Reference: observability, SDK resources reference, proxy reference, environment variables, and limits.