Tools & Sandbox

Built-in tools, sandbox posture, and write policy.

BudgetAI tools run inside a layered sandbox with policy and approval controls. Blocked output is discarded, and tool results are escaped so they cannot trick the model into unsafe actions.

Built-in tools

Read, write, and edit files; run shell commands and background tasks; use git; search the web and fetch pages safely; create documents and media; and generate images.

Sandbox postures

PostureBehavior
StandardSensible defaults for local work
StrictTighter limits on tool actions
LockedMinimal surface; most actions need approval

On Linux, Bubblewrap isolates running code; on macOS, sandbox-exec (Seatbelt) does the same. If BudgetAI is denied the same action repeatedly, it pauses itself instead of retrying.

Write policy

File writes are contained to the agent's workspace by default. Attempts to write outside it require explicit permission.

Do not weaken sandbox layers or approval gates to make something "just work." They exist to keep tool use safe.