What it does
Five ideas. Everything else follows.
kosa8 is a container platform where the unit is a whole machine you can save, copy and put back. Once that is true, a lot of things that were hard become one command.
The five
Each one is a command, and each command is on this page working.
Save a computer mid-thought.
Not the disk — the whole machine, including memory and every running process. A checkpoint takes about seven tenths of a second and you can take one whenever you are about to do something you might regret.
Checkpoints →kosa8 checkpoint sealed --label "before the risky bit" checkpoint sealed-ckpt-1 (before the risky bit) saved in 697ms rewind with `kosa8 rewind sealed`
Put it back, still running.
Rewinding is not a restart. The processes resume from the instant you saved them, which is why it takes half a second instead of a boot.
kosa8 rewind sealed sealed is back at sealed-ckpt-1 (485ms) processes resumed from just now
Try every idea at the same time.
Fork the machine once per approach. They all start from the same instant, run in parallel, and the ones that failed are cleaned up for you.
Forking →kosa8 explore sealed --variant 'echo one' --variant 'false' --variant 'echo three' forked 3 branches in 3258ms BRANCH RESULT TIME VARIANT sealed-explore-712651 ok 275ms echo one sealed-explore-712652 error 3ms false *sealed-explore-712653 ok 272ms echo three winner: sealed-explore-712653
Give each assistant its own machine.
Its own kernel, its own network, and an allow-list you write. A tool that tries anywhere else does not get a polite refusal — it gets a dead socket.
Sandboxes →kosa8 sandbox create sealed --egress-allow api.anthropic.com sandbox sealed ready in 584ms kosa8 sandbox run sealed alpine sh -c 'wget -qO- https://example.com' example.com: refused ↑ only api.anthropic.com was allowed out
Run the model on your own machine.
Pull a model from Hugging Face and serve it locally, on both the OpenAI and the Claude API. Nothing leaves the laptop and nothing is billed per token.
Local models →kosa8 model serve qwen2.5-0.5b-instruct-q4_k_m kosa8 model gateway on http://127.0.0.1:7434 OpenAI: POST /v1/chat/completions Anthropic: POST /v1/messages {"role":"assistant","content":"Kosa8 works."} 175 tok/s · $0.00 · 0 bytes egress