Tarit offers a fast self-hosted sandbox for AI agents
Tarit is a self-hosted sandbox system for running AI agents and environments safely. Its core is a hypervisor that starts and manages small quickly. It can be used instead of Firecracker, but it is designed around AI workloads rather than mainly serverless computing.
One highlighted difference is live snapshots, which can save the state of a without stopping it. Tarit also includes an orchestrator that places across servers, creates highly available clusters, keeps a warm pool of ready machines, and handles networking and monitoring. On a bare-metal server, its benchmark shows 35 milliseconds at p99 to get a from the pool and run code, and about 80 milliseconds to resume a small sandbox from a snapshot.
It can be self-hosted on cloud machines with enabled, or preferably on bare metal.
Key points
- Tarit is a self-hosted sandbox system for AI agents and environments.
- It is presented as an alternative to Firecracker, with AI-focused features such as live snapshots.
- The orchestrator manages placement, clusters, warm pools, networking, and monitoring.
- The stated benchmark is 35 milliseconds at p99 to get a and run code from a ready pool.
- It can run on cloud machines with or on bare-metal servers.