Sandbox Agent tools (BETA)
Inside the sandbox, the agent doesn't just reply — it calls tools to act. During a turn, every tool call shows up live in the UI. This page describes the tools available in the Sandbox Agent BETA feature (overview).
Files
write_file / read_file
Write a file into the sandbox workspace, and read it back. This is how the agent produces code, data, or a deliverable to later publish as an artifact.
filesystem
Explore the workspace: list a directory's contents and get file metadata (stat). Useful for the agent to get its bearings before acting.
Execution
run_code
Run a code snippet (in the agent's runtime, Python or Node) and get its output. Ideal for a one-off computation or a quick check.
terminal
Run shell commands in the sandbox. The agent can install an available dependency, manipulate files, run tests, and so on.
python_context
A stateful Python namespace that persists across calls within a single sandbox: variables and imports you define stay available from one call to the next, like a notebook. (Reminder: this state is gone when the sandbox expires, ~10 min.)
Persistence and deliverables
checkpoint
Create, list, restore, and delete checkpoints of the filesystem state. This is how the agent keeps an intermediate state and returns to it mid-task.
publish_artifact
Publish a workspace file as a downloadable artifact (with an expiry). See workspace and artifacts.
Web
web_search
Web search via Brave. May be gated: this tool isn't necessarily available on every beta account.
web_fetch
Fetch the contents of an authorized public page. Access is SSRF-guarded (only allowed targets are reachable) and the tool may ask for confirmation before fetching a page. May be gated depending on your account.
Next steps
- Conversations and turns — watch tool calls stream in.
- Workspace and artifacts — upload files, publish deliverables.