1
1 line
The line you write.
Adding a field is one declaration — a name and a type. The database, API, screens and agent tools all come from that line, so there is nothing else to write by hand.
The TypeScript framework, agents included.
Powered by Bun
- Web
- App
- Server · DB
- Infra
- Agent
New · Akan.js v3 — agents join the full stack
Build a screen.Agents can use it.Build a server.Any AI can run it.
No tool schemas, no MCP server to write, no second permission model. The app you build for people is already the one AI can use — on the model you choose.
Paste it into Claude Code or Codex, and it sets up the workspace and starts your app. See the prompt or the terminal command ↓
1 line×8 layers×6 platforms×people & agentsThe usual way
Your next user doesn't click.
Agents already read screens and call APIs for the people they work for. Getting an app ready for them is usually a second project — stacked on the first one you already wire by hand.
▾+ namethe first project
- schemaDefine the DB schema
- queryAdd the query field
- serviceAdd the service logic
- apiAdd the API field
- fetchAdd the fetch field
- typeDeclare the client type
- stateDeclare the state management
- uiDeclare the UI prop
…for one field. Then again on every platform you ship.
Screen → agent
IcecreamOrder.Zone.tsx
Server → AI
Your server is already an MCP server.
01 / 04
Point any MCP client at your app.
/mcp is on by default. Every endpoint whose guards admit the caller is a tool, described by the dictionary you already write.
icecreamOrder.signal.ts02 / 04
Sign-in and consent happen on your app.
OAuth 2.1 ships with libs/shared. The AI gets a token for this one user — exactly their rights, revocable at any time.
03 / 04
Then it just works.
Ask in plain words. It calls your endpoints through the same guards and services as your screens — and the board that's open updates live.
04 / 04
And it can't do what it shouldn't.
refundIcecreamOrder is guarded by Person, so it never reaches the shelf. To the AI it looks exactly like a tool that doesn't exist.
icecreamOrder.signal.ts
Guard → everyone
One rule. Three kinds of users.
You write a guard once per endpoint. It decides for a person on the screen, for the agent in their tab, and for an AI calling over MCP.
| Action | PersonPerson | AgentIn-page agent | MCPAI over MCP |
|---|---|---|---|
Place an order createIcecreamOrder | ✓ | asks first | ✓ |
Serve an order serveIcecreamOrder | ✓ | ✓ | ✓ |
See today's sales icecreamOrderSummary | ✓ | ✓ | ✓ |
Refund an order refundIcecreamOrder | ✓ | ✓ | — not on the shelf |
Remove an order removeIcecreamOrder | ✓ | asks first | ✓ |
asks first — the in-page tool waits on an approval card · not on the shelf — the endpoint never reaches MCP
Refusals give nothing away
A tool an agent may not use answers exactly like one that doesn't exist, so the shelf leaks nothing.
Rate-limited per caller
MCP calls are capped at 120 a minute and 8 at once for each caller.
Secrets stay home
Hidden and secret fields are stripped before anything reaches a model.
Connections end when you say
Revoke a connection and its next call is refused. The chat relay keeps no session and no transcript.
How it works
All of it is one line.
The screen an agent drives, the server any AI calls, the guard that judges them all — none of it is a second project. It all grows from a line like this.
1name: field(String)
Scroll and watch it grow
×8
8 layers
Through every layer.
That one line drops through the schema, query, service, API, fetch, client type, state and UI prop — the eight places of the first project, now changing together. Nothing to chase, nothing to miss.
×6
6 platforms
Onto every platform.
The same code ships as SEO-ready web, iOS and Android apps, and macOS, Windows and Linux desktop apps — with native-level screen transitions, not a wrapped website. One implementation to maintain, not six to keep in step.
×2
people & agents
For people and agents alike.
Every guarded endpoint becomes an MCP tool and every control on screen an in-page agent tool, behind the same guards people pass. The screens, servers and guards you saw above all start here — in code you already wrote.
1 line × 8 layers × 6 platforms × people & agents
One star. The whole sky.
One line runs through every layer, lands on every platform and reaches everyone who uses it — people and agents. Type-safe from the database to the screen, and all you wrote was the line.
Agents build it, too
AI coding turns to spaghetti past a certain size.
The faster an agent writes code, the more file paths, names, structures, and declaration styles drift apart — until review and maintenance fall over. Akan stops this at the source with strict rules.
This is what we mean by agentic full-stack.
It runs in both directions. Agents use the app through the same guards people pass. And agents build it on strict rules and fixed blocks — upload, login, admin, chat, boards, alerts — so they produce nothing but consistent code. Not an abstract idea, but quality that rules make.
This site
These docs run on it, too.
Ask the docs
Every docs page carries the in-page agent. Ask about a topic; it searches the docs and opens the page for you.
Open the docs26MB → 8.1MB
Client build output in v3
3.5ms → 0.9ms
Hydrating 1,000 rows on the client
−33%
Time for a 50-row list query
v3 benchmark
Startup 2× faster, a third less memory
The TypeScript framework, agents included.
Start with one line.Agents included.
Hand one prompt to your coding agent, or run one command yourself, and the workspace is ready. The next line you write ships to web, iOS, Android, desktop, your server and database — and to every agent your users talk to.
Set up a new Akan.js workspace for me (https://akanjs.com). 1. Check that Bun 1.4 or newer is installed (`bun --version`). If it isn't, install or upgrade it as https://bun.sh describes. 2. Ask me for a workspace name and a first app name, short and lowercase, unless I wrote them below. 3. In this directory, run `bunx create-akan-workspace@latest <workspace> --app <app>`. It installs the akan CLI globally, creates ./<workspace> and installs its dependencies. 4. Inside ./<workspace>, start the dev server in the background with `akan start <app>` and check that the URL it prints (http://localhost:8282 by default) loads. 5. Tell me it's running, and that I should reopen you inside ./<workspace>: the project's Akan MCP server and AGENTS.md rules load from there.
