Building by chat
How you build your system in TulipFarm — you describe what you want, and agents build it.
You do not configure TulipFarm by editing files. You describe what you want in chat, and agents build it for you — resource types, agents, skills, and your day-to-day records and knowledge. The files in the soul are the result of those chats, not something you normally write by hand.
The assistant is your front desk
Normal chat answers questions, reads your resources and knowledge, and handles everyday data work directly — creating records, writing knowledge pages, searching. If your intent is clear, it acts; it does not ask for confirmation it does not need.
What the assistant does not do is build your system itself. The moment you ask for something that changes what your instance is — a new resource type, an agent, a skill, or first-time setup — it hands the chat to a specialist.
Building in the same chat
When you ask to build or change your system, the same chat loads the appropriate forge, asks for the details it needs, and builds the artifact.
A request like any of these triggers the handoff on the first turn:
- "Create an invoices resource" · "Track support tickets" · "Store our customers"
- "Build me a CRM" · "Set up a hiring pipeline"
- "Create a support agent" · "Make a skill that drafts replies"
- "Set up my business" (first-time onboarding)
You do not phrase these in any special way and you do not write schema. You say what you want in plain language; the assistant asks the few questions it needs and does the rest.
Forges guide the build
The assistant does not improvise. For each kind of artifact it loads a built-in forge — a structured interview that walks from intent to a finished artifact:
| Forge | Builds |
|---|---|
resource-forge | a resource type and its schema |
agent-forge | an agent |
skill-forge | a skill |
onboarding | a whole starter setup, in one guided session |
Forges ship with TulipFarm. They are not something you install or edit — they are how building works.
The result lives in the soul
When the assistant finishes, the artifact is written to your
soul repository and committed — a resource type becomes
resources/{name}/schema.yml, an agent becomes agents/{name}/AGENT.md. There is no
approval step for soul writes, and the change takes effect immediately — a new resource
type's table is ready without a restart. The git history is your audit trail of every
build.
You can edit those files by hand if you want to — they are plain text in a git repo — but that is the power-user path, not the everyday one. Each concept page shows the underlying format for readers who want it.
What you still do in Settings
Two things are deliberately kept out of chat, because they are operator decisions, not building:
Agents cannot set secrets or change your LLM providers. You configure those under Settings as an admin — see Configure LLM providers and Store a secret. Everything else, you build by chatting.
Where to go next
- Define a resource type — the build chat, step by step.
- Resources, Agents, Skills — what you can build, and the format each produces.
- Soul — where everything you build is stored and versioned.