I build complex AI systems that keep working in production.

Set limits for every agent run.

An open-source Linux sandbox for coding agents and command-line tools. Choose the files, network hosts, and runtime each task can use, then inspect a receipt of the enforced policy, outcome, and observation gaps.

• File boundaries
• Network controls
• Runtime limits
• Run receipts

Platforms

Linux · macOS offline VM preview

Profiles

Claude Code · Codex · OpenCode (experimental)

Records

Run receipt · event journal

Explore Ukai ↗

{
  "phase": "settled",
  "containment": "enforced",
  "outcome": { "kind": "exited", "code": 0 },
  "applied": {
    "network": { "mode": "none" }
  },
  "lifetime": {
    "tree_empty": true,
    "integrity": "verified"
  }
}