Which package do I need?

A 5-question decision tree from your use case to the one or two AgentsKit packages you actually need to install.

You don't need all 13+ packages. Most projects use 2–4. Walk through the questions below, install only what each branch tells you to, and skip the rest.

#1. What surface are you building?

Surface	Install
Web chat UI	`@agentskit/react`
Terminal chat UI	`@agentskit/ink`
Server-side / API route / cron / queue worker (no UI)	`@agentskit/runtime`
Multi-agent orchestration (planner + delegates)	`@agentskit/runtime` + `@agentskit/skills`
Authoring a new tool / skill / adapter for distribution	`@agentskit/templates`

Every framework binding (React, Ink, Vue, Svelte, Solid, Angular, React Native) uses the same controller from @agentskit/core underneath, so swapping surfaces later is a one-line import change, not a rewrite.

#2. Which provider talks to the model?

You want…	Install + import
OpenAI / GPT-4o / o1	`@agentskit/adapters` → `openai`
Anthropic / Claude	`@agentskit/adapters` → `anthropic`
Gemini / Vertex AI	`@agentskit/adapters` → `gemini` or `vertex`
AWS Bedrock	`@agentskit/adapters` → `bedrock` (optional `@aws-sdk/client-bedrock-runtime`)
Azure OpenAI	`@agentskit/adapters` → `azureOpenAI`
Local / offline (free)	`@agentskit/adapters` → `ollama`
One key, many models	`@agentskit/adapters` → `openrouter`
Already on LangChain / Vercel AI SDK	`@agentskit/adapters` → `langchain` / `vercelAI` (no rewrite)
Don't decide — try a few	`@agentskit/adapters` → `createRouter` / `createFallbackAdapter`

If you're stuck choosing between adapters, see the capability matrix.

#3. Does the agent need real-world capability?

Need	Install + import
Web search, fetchURL, filesystem, shell	`@agentskit/tools` (built-ins)
GitHub / Linear / Slack / Stripe / Postgres / S3 / 20+ more	`@agentskit/tools/integrations`
Sandbox dangerous code (E2B / WebContainer)	`@agentskit/sandbox`
Wrap any third-party SDK as a tool	`@agentskit/tools` → `defineZodTool` or `defineTool` from `@agentskit/core`
Bridge to Model Context Protocol servers	`@agentskit/tools/mcp`

If the answer is "no tools, just chat" — skip this step. @agentskit/core + an adapter is enough.

#4. Does it need to remember anything?

What to remember	Install + import
Conversation history (single session)	nothing — the controller buffers in-memory by default
Conversation history (across sessions)	`@agentskit/memory` → `fileChatMemory` / `sqliteChatMemory` / `tursoChatMemory` / `redisChatMemory`
Long-term knowledge / RAG	`@agentskit/rag` + a vector store from `@agentskit/memory` (`pgvector` / `pinecone` / `qdrant` / `chroma` / `supabaseVectorStore` / `weaviateVectorStore` / `milvusVectorStore` / `mongoAtlasVectorStore` / `upstashVector`)
Per-user personalization	`@agentskit/memory` → `createInMemoryPersonalization`
Knowledge graph	`@agentskit/memory` → `createInMemoryGraph`
End-to-end encrypted history	`@agentskit/memory` → `createEncryptedMemory`

#5. Does it need ops / observability?

Concern	Install + import
Watching agent runs in dev	`@agentskit/observability` → `consoleLogger`
Per-run cost ceiling	`@agentskit/observability` → `costGuard`
Per-tenant SaaS budgets	`@agentskit/observability` → `multiTenantCostGuard`
LangSmith / OTel / Datadog / Axiom / New Relic	`@agentskit/observability` → matching sink
Audit log + signed trace export	`@agentskit/observability` → `createSignedAuditLog`
Live evals / regression tests	`@agentskit/eval`

If you don't need any of these yet, skip the package entirely. It is purely additive.

#TL;DR

The smallest working agent is two packages:

npm install @agentskit/react @agentskit/adapters

Add @agentskit/tools when the agent needs to do things, @agentskit/memory when it needs to remember, @agentskit/observability when it goes to production.

#Faster path

If even five questions felt like too many, run agentskit init — the interactive picker walks you through these decisions and writes the right package.json for you.

pnpm dlx agentskit init

Architecture at a glance
Concepts — the six contracts every package implements.
Choosing an adapter

Explore nearby

✎ Edit this page on GitHub·Found a problem? Open an issue →·How to contribute →