Edge bundle (Cloudflare Workers / Deno Deploy)

AgentsKit is built so the production hot path on edge runtimes stays under 50 KB gzipped. The trick: every package is independently importable and tree-shakeable, and the only mandatory dep on the chat path is @agentskit/core + one adapter.

#The minimum viable edge bundle

// app/api/chat/route.ts (Vercel Edge / CF Workers / Deno Deploy)
import { openai } from '@agentskit/adapters'

export const runtime = 'edge'

export async function POST(request: Request) {
  const { messages } = await request.json()
  const adapter = openai({ apiKey: process.env.OPENAI_API_KEY!, model: 'gpt-4o-mini' })

  const source = adapter.createSource({ messages })
  const stream = new ReadableStream<Uint8Array>({
    async start(controller) {
      const encoder = new TextEncoder()
      for await (const chunk of source.stream()) {
        if (chunk.type === 'text') controller.enqueue(encoder.encode(chunk.content))
      }
      controller.close()
    },
  })

  return new Response(stream, { headers: { 'content-type': 'text/plain' } })
}

Bundle: @agentskit/adapters (with only openai imported) tree-shakes to ~12 KB gzipped on edge runtimes — well below the 50 KB ceiling.

#Sizing budget

pnpm size runs the full size-limit suite. Per-package limits live in .size-limit.json.

#What to skip on edge

These packages bundle DOM / Node primitives that don't fit edge runtimes:

• @agentskit/react / @agentskit/ink — UI bindings; client-side only.
• @agentskit/sandbox — wraps E2B / WebContainer; needs Node primitives.
• @agentskit/observability — Node-flavored sinks (audit log, devtools server, OTel exporters); the SaaS HTTP sinks (datadogSink, axiomSink, newRelicSink) DO work on edge.
• @agentskit/cli — Node CLI; not for runtime use.

#Per-runtime notes

#Cloudflare Workers

compatibility_flags = ["nodejs_compat"] if your code touches node:crypto or node:buffer. Use the cloudflare-workers starter (agentskit init --template cloudflare-workers) for the full setup.

#Deno Deploy

npm:@agentskit/adapters works as-is. Use Deno.env.get('OPENAI_API_KEY') for provider keys. Use the deno-deploy starter (agentskit init --template deno-deploy).

#Vercel Edge

export const runtime = 'edge' on the route handler. The nextjs starter ships an Edge route at app/api/chat/route.ts.

#Bun

Bun runs the Node hot path directly — no edge constraints. Use the bun starter.

#What runs natively in browsers

@agentskit/adapters ships a webllm adapter that runs LLMs on-device via WebGPU (no inference network call):

import { webllm } from '@agentskit/adapters'

const adapter = webllm({ model: 'Llama-3.1-8B-Instruct-q4f16_1-MLC' })
// Inference: 100% client-side. No API key, no token cost.

@mlc-ai/web-llm is an optional peer dep (~1 MB initial wasm download per model). Pair with @agentskit/react for a fully client-only chat that never hits a server.

#Related

• Cloudflare Workers starter (--template cloudflare-workers)
• Deno Deploy starter (--template deno-deploy)
• Choosing an adapter — which fits edge.