A 10 KB core budget. 24 focused packages. Zero lock-in. Six formal contracts that make every adapter, tool, skill, memory, retriever, and runtime substitutable.
You started building an AI agent last week. You're three libraries deep, two of them fight each other, and nothing you wrote is reusable. This is for you.
⭐ If this saves you from gluing five libraries together, star the repo. AgentsKit is solo-built — a star is the cheapest signal that it's worth continuing, and it's what puts it in front of the next person.
Verified proof
The current evidence ledger, generated from repository sources by scripts/compute-stats.mjs, verifies 24 published packages, 7 framework bindings, 25 native adapters, 50 integrations, 140 catalog providers, 5,000+ catalog models, 21 ready-made skills, 17 memory backends, and 69 recipes.
The core has zero runtime dependencies and a CI-enforced 10 KB gzipped budget.
Every public package has an explicit stability tier, test floor, README, human guide, and agent handoff.
Numeric claims fail CI when their repository derivation, committed snapshot, or public evidence drifts.
Building a real AI agent in JavaScript today means cobbling together five libraries that don't fit. Vercel AI SDK is a beautiful chat SDK with no runtime. LangChain.js drags in 200MB and leaks abstractions at every layer. MCP solves tool interop and nothing else. assistant-ui has 53 components and no opinion about how to compose them.
AgentsKit is the missing kit: small, contracted, composable. Start with one package, grow into a full stack, and stay in plain JavaScript the entire time.
AgentsKit is not only a package. It is the open-source foundation of a full agent ecosystem:
build from composable parts, start faster with ready agents, follow production patterns, and run the result with enterprise controls.
Layer
Go to
What it unlocks
AgentsKit
This repo
The agent building blocks: adapters, runtime, tools, memory, RAG, UI, evals, observability, and MCP
Field-tested patterns for designing, evaluating, securing, and operating agents
flowchart LR
Playbook["Playbook<br/>best practices"]
Registry["Registry<br/>ready agents"]
AK["AgentsKit<br/>composable JS packages"]
Playbook --> AK
Registry --> AK
AK --> Registry
import type { AdapterFactory } from '@agentskit/core'
import { createRuntime } from '@agentskit/runtime'
const localAdapter: AdapterFactory = {
createSource(request) {
const task = request.messages.at(-1)?.content ?? 'your task'
return {
async *stream() {
yield {
type: 'text' as const,
content: `Agent ready. I received: ${task}`,
}
yield { type: 'done' as const }
},
abort() {},
}
},
}
async function main() {
const runtime = createRuntime({ adapter: localAdapter })
const result = await runtime.run('Plan my first production agent')
console.log(result.content)
}
void main()
npx tsx agent.ts
It prints Agent ready. I received: Plan my first production agent. No account, API key, or network request is required. The executable fixture is byte-synchronized with this README and runs in CI; connect any supported provider later through the same adapter contract.
Before and after
Before — the typical "JS agent" stack:
// Pick your favorite: LangChain, raw fetch, Vercel AI SDK + custom runtime,
// MCP client + custom UI, manual ReAct loop, hand-rolled streaming...
// Then wire memory. Then wire tools. Then wire delegation. Then debug.
After — AgentsKit:
import { createRuntime } from '@agentskit/runtime'
import { openai } from '@agentskit/adapters'
import { webSearch, filesystem } from '@agentskit/tools'
const runtime = createRuntime({
adapter: openai({ apiKey: KEY, model: 'gpt-4o' }),
tools: [webSearch(), ...filesystem({ basePath: './workspace' })],
})
const result = await runtime.run('Research the top 3 AI frameworks and save a summary')
That's an autonomous agent. With a tool registry. With memory. With observability hooks. Two imports, six lines.
Swap providers in one line — every other line stays the same:
You only need a single OpenAI streaming call. Use the openai SDK directly — AgentsKit is overkill.
You're shipping a chat SDK to consumers, not an agent. Vercel AI SDK is purpose-built for that and excellent.
You need Python. AgentsKit is JavaScript-first by design. Use a Python framework.
You require enterprise-grade observability today. AgentsKit's observability layer is good but young; LangSmith/Arize/Helicone are more mature integrations right now.
You need every package frozen today.@agentskit/core is v1.0.0, but the rest of the ecosystem is still graduating package-by-package.
Full, honest head-to-head with LangChain.js, Vercel AI SDK, Mastra, LlamaIndex.js, and assistant-ui → AgentsKit vs alternatives.
The packages
Pick what you need. Every package works alone. Combinations work without glue code.
The planner decomposes the task. The researcher and coder execute their parts. Delegation happens through a tool the model already knows how to call — no special syntax to learn.
The same useChat mental model. Real keyboard input. Real streaming. Real tools.
For AI agents reading this
The full public API fits in under 2,000 tokens. Paste the agent-friendly reference into your LLM context and start generating real AgentsKit code immediately. We treat agents as first-class consumers of our docs.
Read these once and you can predict how every package behaves.
Maturity and compatibility
@agentskit/core is at v1.12.8 — API frozen at the minor level, deprecations carry a cycle, contracts pinned to ADRs. The rest of the ecosystem ships on independent beta tracks with explicit stability tiers.
AgentsKit targets Node.js 20+ and modern JavaScript runtimes. Packages ship strict TypeScript declarations and dual ESM/CJS output unless their platform binding documents a narrower target. Browser, edge, React Native, Angular, Vue, Svelte, Solid, React, and terminal compatibility is package-specific and declared in each package README and guide.
The current verified release surface is:
Stable:@agentskit/core, with six formal contracts pinned to ADRs 0001–0006.
Beta: every other package graduates independently under the published stability policy.
Evidence: current package, adapter, integration, provider, skill, memory, and recipe counts come from the generated claims ledger, not this prose.