AgentsKit.js
The TypeScript foundation for building AI agents — runtime, tools, memory, RAG, adapters, headless UI bindings, and production guardrails.
Agents shouldn't be a monolith. AgentsKit.js is a family of small, plug-and-play packages covering the whole agent lifecycle in JavaScript: autonomous runtimes, tools, skills, memory, RAG, headless UI bindings, observability, evaluation, and sandboxing. Install what you need. Everything else stays out of your bundle.
#Start here
Pick a path by what you're shipping.
- Want the real ecosystem in one guide? → Build your first agent
- Want to start from an outcome? → Use cases
- Building a versioned multi-surface chat app? → AgentsKit Chat (application layer on this foundation)
- Just need a headless
useChatbinding? → Quick start · React · Vue · Svelte · Solid · Angular · React Native · Ink - Building an autonomous agent? → Runtime · Tools · Skills · Multi-agent topologies
- Need RAG? → createRAG · Loaders · Reranking
- Shipping to production? → Observability · Security · Evals · Durable execution
#Build path
If you are evaluating AgentsKit, this is the shortest path to understanding the product:
- Build your first agent — see provider + runtime + tools + memory + observability working together.
- Architecture at a glance — understand which layer owns what.
- Production — add observability, security, evals, and CLI workflows.
#Popular builds
- Support agent — customer support with memory, tools, and escalation.
- Research agent — web research, structured summaries, and background jobs.
- Code agent — repository-aware automation with safe tool use.
- Internal copilot — company knowledge with RAG and production controls.
#The substrate
Six stable contracts. Every package is an implementation of one.
| Contract | Role | Deep dive |
|---|---|---|
| Adapter | LLM provider seam | concept · implementations |
| Tool | Function the model calls | concept · implementations |
| Skill | Declarative persona | concept · implementations |
| Memory | Chat history + vector state | concept · implementations |
| Retriever | Context fetching | concept · implementations |
| Runtime | The loop that composes them all | concept · implementations |
Learn the six once; swap implementations forever. The mental model walks through how they compose.
#What's inside
- 22 packages under
@agentskit/*— foundation, providers, UI bindings, runtime, capabilities, observability, infrastructure. Full index. - 7 framework bindings sharing one
useChatcontract. UI matrix. - 25 native adapters plus a catalog of 140 providers. Providers.
- 50 integrations for tools agents call. Integrations.
- 60+ recipes grouped by theme. Recipes.
- Open specs — A2A, Manifest, Eval Format, AgentSchema — portable JSON. Specs.
#Keep reading
- Comparison — AgentsKit vs LangChain, Vercel AI, Mastra, LlamaIndex.
- Migrating — port from Vercel AI SDK, LangChain.js, Mastra.
- For agents — dense LLM-friendly reference per contract.
- Examples — live interactive demos.
- Contribute — built in the open. Discord is optional; product chrome does not promote it.
- AgentsKit Chat — versioned chat applications on this foundation.