The animation shows Kinu trying several approaches to a task and keeping the one that passes.
Give Kinu a job and close the laptop. It keeps working in the cloud, uses the apps you connect, and asks before risky commands. Run it on kinu.run, on your own Cloudflare account, or in your terminal.
curl -fsSL 'https://kinu.run/install.sh' | bashPlan mode. Kinu reads and plans, but changes nothing until you approve the plan or mark the lines to change.
Live apps. Kinu can build a small web app, like this dashboard, and open it beside the chat. The app gets only the files and tools you give it, never your keys.
Same agent either way. You choose when you create a workspace.
Cloud agents run on Cloudflare. Start one from chat, on a schedule, or from a webhook. When a job needs a full computer, the agent gets one of its own, with a preview link you can open.
Run kinu connect on your Mac or Linux machine, and agents can use it in the workspaces you allow.
Want it on your own Cloudflare account? Sign in at kinu.run/deploy. It needs Workers Paid, $5 a month. Or follow the self-hosting guide.
No account needed. Bring your own API key, or use Workers AI once you sign in.
curl -fsSL 'https://kinu.run/install.sh' | bash
kinu create workshop --mode local
kinu chat workshopkinu exec runs one task with no prompts and exits with the result.
kinu exec -w workshop --json "Review this diff and report what you checked."The step fails if the task errors or reaches for access you didn't grant ahead of time. Your tests and reviewers still decide what merges.
For a cloud workspace, set KINU_TOKEN from kinu tokens create.
Rate a turn or correct it, and Kinu adjusts. Its edits land in a changelog, and you can undo them.
It writes tools as it goes. The ones that help rise to the top, and the ones it stops using retire.
Every few rated turns, it reviews them and makes small, tested edits to its prompt, memory and skills.
When a problem keeps coming back, it writes a new loop, replays your recent turns through it, and rolls back if it does worse.
A swarm puts several agents on one problem, each trying something different. Give it a command that scores the result, and it builds on whatever scores best.
The task, and a command that scores the result.
Change the algorithm
Change the data structure
Split the work differently
Kinu keeps every score and builds on the approaches that improved.
The agent, both backends and the CLI are MIT-licensed.