TypeScript framework · MIT licensed
Build agentic software for people and agents.
Describe your models once. smrt gives you the screens people use and the tools agents use, and makes both play by the same rules.
npx -y @happyvertical/smrt-cli@0.62.10 kitchen my-app
Opens the planner in your browser. Pick your features, then send them to the kitchen.
src/lib/objects/Widget.ts
@smrt({ api: true, mcp: true, cli: true })
export class Widget extends SmrtObject {
name = '';
inStock = 0;
leftHanded = false;
}Showing what a person sees
| Name | Kind | In stock |
|---|---|---|
| Left-handed doodad | left-handed | 40 |
| Gizmo, regular | 400 | |
| Nick-nack, assorted | 12 |
→ tool call widget_list
← result
[
{ name: "Left-handed doodad", inStock: 40, leftHanded: true },
{ name: "Gizmo, regular", inStock: 400, leftHanded: false },
{ name: "Nick-nack, assorted", inStock: 12, leftHanded: false }
]Same widgets. Same rules. One permission for both: widgets.read
Agentic features, out of the box.
Here is what agents can do inside a smrt app, and how you stay in charge of them. An agent works as a user, with that user’s permissions and nothing more.
- An assistant on the page A chat dock that works with the page you’re on, through that page’s own tools. Anything destructive waits for a person.
- Page tools In a browser that supports page tools (WebMCP), an agent works with the page as the signed-in person. Read-only until you allow more, and a person applies any form change.
- Agents on a schedule Set an agent to run on a schedule. It works as the user it is bound to, with that user’s rights and only the tools you list.
- Approvals Decide who has to say yes, and keep a record of every answer. The assistant’s destructive requests wait for a yes too. Each yes runs that one request, once, as the person who asked.
- Reports on request Describe a report, or let an agent draft one. It is saved only after a person confirms, and each run shows only what the person running it may see.
- AI usage Every model call leaves a row: which prompt asked, which model answered, the tokens, an estimated cost and the time. One command turns the rows into a report.
- Choose your models Your code asks for a quality level, called a tier, from cheap to best. You pick the AI model behind each one. Add one package, smrt-inference, and a call that fails on a quota, a bad key or a missing model moves on to the next model in your list.
- Incoming files and email Each one is kept as evidence, read when you ask, and turned into suggestions. A reviewer decides. This ships as a preview, with automation off.
Also in the box: agent memory, search by meaning, voice on the device.
The parts every app needs are already built.
Then there are the everyday parts: customers, invoices, timesheets, chat. Switch on what you need and save your energy for the part that is actually yours.
- Sell Customers, Orders, Invoices, Products, Quotes, Subscriptions, Payments
- Buy Vendors, Purchase orders, Expenses, Inventory
- Make Assemblies, Production, Stock
- Team Employees, Timesheets, Qualifications, Approvals, Roles
- Publish Articles, Images, Video, Voice, Comments, Social
- Talk Chat, Email, Slack, Support
- Grow Campaigns, Analytics, Reports, Ads, Affiliates
- Plan Projects, Tasks, Events, Places
- Assistant Agents, Tools, Memory
One model, every way in.
You describe each thing in your app once, as a TypeScript class called a model. smrt reads it and generates the rest, so the screen a person uses and the tool an agent calls come from the same model.
- A tableThe fields tell smrt what to store.
- An APIRoutes and clients for exactly the actions you allow.
- Agent toolsThe same actions, described so an agent knows when to use them.
- CommandsFor the terminal and for scripts.
- FormsFields and labels, straight from the model.
- PermissionsHowever a request arrives, the same checks apply.
Three ways to a running app.
They all end in the same place: a real project on your machine.
- in the browser
Plan it first
Tick the features you want in build-a-boilerplate, poke at a working mock-up, and export it as a cookbook.
Open the planner → - one command
Start from a cookbook
A cookbook is a short file naming what your app is made of. The CLI turns it into a project. A bakery and a yoga studio are included.
Cook one → - from scratch
Clone the starter
The template on its own: one owner, a local database, a notes model and an assistant. Add things as you go.
Get smrt-start →
Agents are welcome here too.
Your coding agent should not have to guess how smrt works. This site is written for it, and for you.
llms.txtA directory of the whole site, one line per page.any page + .mdEvery page has a Markdown copy at its address plus .md.- Development MCP Read-only knowledge of your workspace and installed packages.
- Application MCP The operations your app permits, offered to an agent as tools.