The right computer for each task
Sandbox templates
Five base profiles, with practical recipes for the work your agent needs to do. Each page covers preparation, a starting resource budget and useful output artifacts.
These profiles document the execution design. Recipes adapt a base profile; they do not add provisioned images. Runtime access and supported image versions depend on your connected execution service. Start with the quickstart.
Base profiles
- python · 2 vCPU / 4 GB starting profilePythonPlan a Python sandbox for code interpreters, pandas analysis and matplotlib charts, with bounded input, useful artifacts and explicit resource limits.View profile →
- node · 2 vCPU / 4 GB starting profileNodePlan a Node and TypeScript sandbox for generated scripts, package tests and app tooling, with pinned dependencies and structured execution results.View profile →
- browser · 2 vCPU / 4 GB starting profileBrowserPlan a Chromium and Playwright sandbox for browser agents: isolate sessions, scope destinations and return screenshots and structured observations.View profile →
- devbox · 4 vCPU / 8 GB starting profileDev boxPlan a multi-language development sandbox for repository work, builds and tests, with scoped checkout access and reviewable patch artifacts.View profile →
- custom · 2 vCPU / 4 GB starting profileCustom imageDesign a custom sandbox image with a pinned toolchain, reproducible build inputs and a smoke test before turning it into a reusable template.View profile →
Workload recipes
- devbox · 4 vCPU / 8 GB starting profileCoding agentA Dev box recipe for agent-driven repository edits: baseline tests, bounded command execution and patches that a human can review before publishing.View recipe →
- devbox · 4 vCPU / 4 GB starting profileGo builds and testsA Go sandbox recipe for compile and test loops, with pinned modules, bounded test parallelism and separate handling for CGO dependencies.View recipe →
- devbox · 4 vCPU / 8 GB starting profileRust builds and testsA Rust sandbox recipe for Cargo builds, tests and native dependencies, with locked versions, bounded build jobs and explicit artifact collection.View recipe →
- python · 2 vCPU / 4 GB starting profileDocument processingA Python recipe for PDF and office-document extraction, with page limits, preserved source references and separate OCR resource planning.View recipe →
- python · 2 vCPU / 4 GB starting profileSQLite analysisA local SQL analysis recipe for uploaded SQLite databases, using read-only inputs, bounded queries and aggregate results for the agent.View recipe →
- node · 2 vCPU / 4 GB starting profileWeb app previewA Node recipe for agent-built web previews, with scoped ports, bounded server lifetimes and a separate production build before review.View recipe →
- browser · 4 vCPU / 8 GB starting profileBrowser QAA Playwright recipe for repeatable web app checks, with deterministic fixtures, failure traces and separate browser contexts for each test.View recipe →
Choose a profile for your first task
Create a workspace, compare the profiles and plan a bounded execution workflow.
gpuOS · where models think
Need the model too? Run it on gpuOS
gpuOS serves open models on your own GPUs behind one OpenAI-compatible API. The model reasons on gpuOS, the agent acts in a cpuOS sandbox.