Choose your path
Three paths, one output
Section titled “Three paths, one output”Every path ends in the same place: an actor generated from the kaboum template. The differences are what you need installed, what you end up writing, and who each path suits.
Decision matrix
Section titled “Decision matrix”| Path | What you need | What you build | Best for |
|---|---|---|---|
AI-assisted (/create-actor skill) |
Claude Code, and read access to a kaboum repo checkout | A full actor: the skill interviews you, writes a discovery spec, renders the scaffold, then implements real domain code — models, client, tools, tests | Kreators who want a complete actor with domain logic and minimal manual boilerplate |
| Copier CLI | Python 3.11–3.13, copier 9.4.0+, Docker (PostgreSQL 14+ and Redis 6+ come via the generated docker-compose.yml) |
The scaffold yourself with copier copy, then hand-write your domain code against the seven extension points |
Developers who want full control of every line, reproducible regeneration, and CI-friendly builds |
| kreator-sandbox (in-app) | A Kaoohi account and a browser — no local environment | A configuration-level tryout: an ephemeral kompanion you can talk to, with connectors in mock mode | Validating an idea before writing anything, and low-code experimentation |
All three are first-class. Nothing about the sandbox makes it a toy path, and nothing about the CLI makes it the “real” one — pick by what you have and what you want to write.
On the Copier CLI path, the experimental ](en/reference/sdk/) can
script the loop for you: kaboum new renders the scaffold, kaboum validate
checks the emitted config, kaboum publish builds and pushes the image.
Which track are you on?
Section titled “Which track are you on?”The paths also split along two tracks:
| Track A — technical developers | Track B — low-code kreators | |
|---|---|---|
| Comfortable with | Python, Docker, terminals, git-hosted repos | Configuration, prompts, the Kaoohi app |
| Natural path | AI-assisted skill or Copier CLI | kreator-sandbox, then Copier CLI with defaults if you need custom code |
| You extend the actor by | Writing domain code (client, models, tools, migrations) | Adjusting configuration, prompts, and declared connectors |
| You verify by | make kaoohi-config, ruff, pytest, local uvicorn |
Talking to the sandbox kompanion and checking its responses |
Track B kreators are not second-class: the template’s extension points that matter most (system prompts, MCP servers via environment variables, config fields) need little or no Python.
Sandbox limits, stated plainly
Section titled “Sandbox limits, stated plainly”The kreator-sandbox is the fastest start, so know its shape before you lean on it:
- Sandboxes are ephemeral kompanions — they are reaped when their time-to-live expires (30 minutes by default).
- At most 2 concurrent sandboxes per kreator.
- Each session has a token budget (100,000 tokens) and restricted network egress.
- Connectors run in mock mode — flows work, but nothing reaches real external services.
Choose by scenario
Section titled “Choose by scenario”- “I have an idea and want to poke at it tonight.” Sandbox. No installs, and the 30-minute sandbox is enough to feel out whether the idea holds.
- “I know exactly what I want and I write Python.” Copier CLI. The scaffold lands in one command and you own every line from minute one.
- “I want a full actor but not the boilerplate.” AI-assisted. The
/create-actorskill interviews you, then writes the domain code with you reviewing. - “I’m not sure which connectors my actor needs.” Sandbox first — mock-mode connectors let you rehearse WhatsApp and n8n flows before you commit to declarations that ship in the actor’s metadata.
- “My actor must cooperate with other actors.” Any path works — mesh
participation is a generation-time option (
fleet_mesh) — but the AI-assisted path is the quickest route to a coherent suite, because the discovery interview covers mesh roles and dependencies before any code exists.
One more tie-breaker: you can switch paths without losing work. The sandbox teaches you what to want; the CLI and the skill both start from the same template, so what you learned transfers — and an actor generated by the skill is the same repo you would have generated by hand.
If in doubt
Section titled “If in doubt”Start in the sandbox. It costs nothing to set up, and everything you learn there (what your actor should do, which connectors it needs, how it should answer) transfers directly to the AI-assisted or CLI path when you need real code. When you outgrow mock connectors or want tests and a repository, follow the ](en/quickstart/).