About

The lab studies what it takes to leave AI agents running unattended: for hours, on real tasks, with access to tools. Most of the effort goes into what surrounds the model: where its code runs, what it can reach, what it remembers, and what it costs.

The work is infrastructure: an agent runtime, a model gateway with cost-first routing, agent memory, and benchmark harnesses. Each piece is built to be used on its own as well as together.

The pieces are also used on each other. The runtime's model calls go through the gateway, and the benchmark runs are where memory, routing and cost limits get tested under load, so each tool is exercised before it is released.

Run it yourself

The tools are designed to be self-hosted on hardware you control, down to a small box at home. Nothing depends on an account with us.

Keep less

You don't have to protect what you do not have. Kasumi (霞, "mist"), our model gateway, strips identifying headers, scrubs personal data from requests, and stores no prompts or responses by default.

Contain by default

Code written by an agent runs in a sandbox with no network access and no secrets. Outbound requests pass through a per-agent allowlist, and sensitive actions wait for a person to approve them.

Measure it

Model spending is metered per project and can be capped. Agents are tested on long, open-ended tasks, because short prompts say little about how an agent holds up over hours.

The code is written for release under the MIT license. Repositories open as each project is ready for other people to run.

Questions about the work, research, or collaboration on open tools: contact@asukalabs.co.