Create a Kiro CLI agent from a screen recording
Record the task once in Ditto AI, review it, and pick Kiro when you run it. Ditto writes a skill with the procedure into your Kiro skills folder and an agent file into your Kiro agents folder, carrying the MCP servers and the tool allowlist it derived. You then start it with kiro-cli chat and the agent's name.
What you need
- A Mac with Apple silicon and macOS 15 or later, with Ditto AI installed.
- Kiro CLI on the same Mac, on your
PATH. Ditto only offers hosts whose command it can find.
Step by step
- Record the task once, across the apps it really uses.
- Review the steps, inputs and tools Ditto found, and fix anything wrong.
- Connect the tools that need your account. Each is proven with one read-only call before it counts as configured.
- Choose Kiro as the host and run it.
- Start it later with
kiro-cli chat --agent <name>, where the name is the agent's.
What Ditto writes into Kiro
| Where | What |
|---|---|
| Your Kiro skills folder, under the agent's name | SKILL.md: the goal, the steps in order, the inputs and the rules |
| Your Kiro agents folder | An agent file with the MCP servers the task needs and the tools it may call |
The agent file carries the same servers and the same allowlist Ditto derives for every host; only the shape differs per host. Access per tool is a lookup in a fixed table, not a decision the model makes, and anything Ditto can't classify counts as a write.
One caveat: Kiro's sandbox and the browser
Kiro runs MCP servers inside its own sandbox, and Chrome's sandbox can't start inside that. So for Kiro only, the headless browser Ditto uses for public web steps runs with its own sandbox off, while Kiro's sandbox still contains it. Other hosts keep Chrome's sandbox on.
If that trade-off matters for a task, run the agent in Ditto itself or in Claude Code instead.
Other hosts
Ditto can also install into Claude Code, or run the agent with its own bundled runner, where you can have every write shown to you as a draft first.