AgentsKit

How AgentsKit fits the landscape

A practical lens — not a performance benchmark. Choose what matches your stack and how much control you need.

This table is wide — scroll sideways on small screens to compare all columns.

Dimension	Vendor-tied or closed UI SDKs	Opinionated web framework paths	AgentsKit
Ownership	Often bound to one vendor console, hosted-only features, or opaque upgrade paths.	Open source with a strong “happy path” for a specific ecosystem (e.g. React streaming patterns).	MIT, modular npm packages — you own deployment; swap adapters without rewriting the UI layer.
UI model	Pre-built widgets; limited theming or escape hatches.	Conventions and examples centered on one stack's idioms.	Headless primitives (`data-ak-*`) plus optional theme — bring MUI, shadcn, or your design system.
Surfaces	Usually web-first; terminal or headless agents are an afterthought.	Web-first; depth outside the browser varies.	Same contracts for React, Ink (terminal), CLI, and headless `runtime`.
Extension	Plugin tiers, closed extension APIs, or vendor-specific tool schemas.	Composable inside the framework's world.	Shared core types — adapters, tools, memory, RAG, skills compose on `@agentskit/core`.

Best fit: one stack across browser, terminal, and headless agents, with control over adapters and tools. Skip if: a single vendor widget is enough, or you are fully committed to one framework's AI path.

Closed SDKs and opinionated paths slow you down

You still own streaming edge cases — unless your stack is built to compose

Vendor-tied or closed layers — upgrade schedules, hosted-only features, and opaque limits box you in when you need custom tools or self-hosted inference.
Highly opinionated frameworks — productive inside their happy path; painful when your product spans web, terminal, and headless workers with one contract.
Fragile hand-rolled streams — partial chunks, retries, and aborts still leak into UX (stuck “typing”, duplicated messages, lost tool results) when glue code diverges per surface.

AgentsKit is MIT-licensed and extension-first: adapters, tools, memory, and RAG sit on shared primitives so you keep control and swap pieces without rewriting the product narrative.

From ~50 lines of glue to ~10 lines of product

Less code to review, test, and regret when the API shape changes next month

Before — ~50 lines

const [messages, setMessages] = useState([])
const [input, setInput] = useState('')
const [streaming, setStreaming] = useState(false)
const abortRef = useRef(null)

const send = async () => {
  setStreaming(true)
  const userMsg = { role: 'user', content: input }
  setMessages(prev => [...prev, userMsg])
  const res = await fetch('/api/chat', {
    method: 'POST',
    body: JSON.stringify({ messages: [...messages, userMsg] }),
    signal: abortRef.current?.signal,
  })
  const reader = res.body.getReader()
  const decoder = new TextDecoder()
  let text = ''
  while (true) {
    const { done, value } = await reader.read()
    if (done) break
    text += decoder.decode(value)
    setMessages(prev => [...prev.slice(0,-1),
      { role: 'assistant', content: text }])
  }
  setStreaming(false)
}
// + error handling, auto-scroll, cleanup...

After — 10 lines with AgentsKit

import { useChat, ChatContainer, Message, InputBar }
  from '@agentskit/react'
import { anthropic } from '@agentskit/adapters'
import '@agentskit/react/theme'

function Chat() {
  const chat = useChat({
    adapter: anthropic({ model: 'claude-sonnet-4-6' })
  })
  return (
    <ChatContainer>
      {chat.messages.map(msg =>
        <Message key={msg.id} message={msg} />
      )}
      <InputBar chat={chat} />
    </ChatContainer>
  )
}

One architecture, every surface

Jump to the guide you need — each link matches an @agentskit/* package on npm.

React Ink CLI Runtime Adapters Memory RAG Tools Skills

Model providers — swap in one line

AnthropicOpenAIVercel AI SDKAny ReadableStream