Kakera

Kakera (欠片) is Japanese for "fragment" or "shard" — each route is its own little fragment, sandboxed and loaded on demand.

A file-based routing framework for Cloudflare Workers. Each route file runs as an independent Dynamic Worker — fully isolated, loaded via the Worker Loader binding.

Status: experimental.

Install

npm i kakera-worker

How it looks

my-app/
  routes/
    index.ts        # GET /
    users.ts        # /users/*
    posts.ts        # /posts/*
  src/
    app.ts          # host worker entry
  build.ts          # bun build script
  wrangler.jsonc
  package.json

Each route is just a worker — typically a Hono app:

// routes/users.ts
import { Hono } from 'hono'

const app = new Hono()
app.get('/', (c) => c.json({ users: [] }))
app.get('/:id', (c) => c.json({ id: c.req.param('id') }))

export default app

The host Worker dispatches by the first path segment (/users/123 → users.ts, then forwards /123). Routes are sandboxed from each other.

Host worker

// src/app.ts
export { app as default } from 'kakera-worker'

That's it. Or with options:

import { kakera } from 'kakera-worker'
export default kakera({ dir: 'subdir' }) // fetches subdir/<name>.js via ASSETS

extensions option (default ['js']):

kakera({ extensions: ['js', 'mjs'] })

Build script

build.ts uses Bun's Glob API so route discovery is shell-independent and works whether you have only .ts, only .tsx, or both:

// build.ts
import { Glob } from 'bun'

const entrypoints = [...new Glob('routes/*.{ts,tsx}').scanSync('.')]

const result = await Bun.build({
  entrypoints,
  outdir: './dist',
  target: 'browser',
  format: 'esm'
})

if (!result.success) {
  for (const log of result.logs) console.error(log)
  process.exit(1)
}

Wrangler config

Routes are pre-bundled per-file by bun build, and the output directory is served via the ASSETS binding. Wrangler's [build] runs the bundler on startup and re-runs it when watch_dir changes — wrangler dev is the only command you need.

// wrangler.jsonc
{
  "name": "my-app",
  "main": "src/app.ts",
  "build": {
    "command": "bun run build",
    "watch_dir": "routes"
  },
  "assets": { "directory": "dist", "binding": "ASSETS" },
  "worker_loaders": [{ "binding": "LOADER" }],
  "compatibility_date": "2026-03-17"
}

Scripts

{
  "scripts": {
    "dev": "wrangler dev",
    "build": "bun run build.ts",
    "deploy": "wrangler deploy"
  }
}

• wrangler dev — runs [build].command (= bun run build), then starts workerd. Edits in routes/ trigger a re-bundle and reload.
• wrangler deploy — same flow, but ships to Cloudflare.

Try the example

cd example
bun install
bun run dev

Why

• Isolation by default. Each route is its own Worker — bugs, deps, and runtime crashes can't leak between routes.
• Tiny host bundle. The host worker is ~1.6 KiB. No runtime bundler shipped.
• Standard tooling. Bundling is bun build. Watch-and-rebuild is Wrangler's [build]. Nothing magic.
• Per-route bindings (planned). Each route can eventually have its own scoped set of bindings.

How it works

References

• Dynamic Workers
• Worker Loader binding
• Wrangler custom builds
• Hono