AgentsKit.js
AgentsKit.js is an open-source TypeScript foundation for building AI agents in JavaScript. Twenty-two MIT-licensed npm packages cover runtime, tools, memory, RAG, observability and headless chat UI, with every layer swappable behind six formal contracts.
What is AgentsKit.js?
AgentsKit.js is an open-source foundation for building AI agents in JavaScript and TypeScript. Its premise is blunt: ship agents without gluing eight libraries together. Instead of one monolithic framework it publishes twenty-two @agentskit/* packages on npm under the MIT licence, and you install only the ones a given problem needs.
Everything rests on six formal contracts — Adapter, Tool, Skill, Memory, Retriever and Runtime — pinned to architecture decision records. The core package that carries them measures roughly 5.17 KB gzipped at version 1.0, with zero dependencies and a size budget enforced in continuous integration. Every layer above it is optional and substitutable, which is the whole argument: swap OpenAI for Claude, React for a terminal interface, or in-memory storage for a vector database, without rewriting the orchestration underneath.
The adapter layer spans 25 native adapters, 140 providers and 5,162 models, local runtimes such as Ollama, LM Studio, vLLM and llama.cpp included. The runtime package executes a ReAct loop and adds planning, speculative execution of parallel candidates, self-debugging from its own traces, durable execution that resumes after a crash, and four multi-agent topologies: supervisor, swarm, hierarchical and blackboard. Tools arrive as 50 verified integrations — Slack, Notion, GitHub, Stripe, Shopify, Postgres, Linear and the long tail — behind one install-configure-execute contract, plus a bidirectional Model Context Protocol bridge that talks to Claude Desktop, Cursor and Windsurf.
Memory splits into conversation history and vector recall across 17 backends, with encrypted, hierarchical, graph and virtualised variants. Retrieval ships as a one-line RAG pipeline with chunking, hybrid vector and BM25 search, reranking and eleven document loaders. On the interface side a single useChat contract drives headless components in React, Vue, Svelte, Solid, Angular, React Native and Ink for the terminal.
Production concerns get their own packages: observability through LangSmith, OpenTelemetry and Langfuse, evaluation suites with deterministic replay and CI reporters, and six security primitives covering PII redaction, prompt-injection detection, rate limiting, signed audit logs, mandatory sandboxing and human-in-the-loop approvals. The project declares itself beta, and only the core package is marked stable.
What it does
- Run an autonomous ReAct agent that observes, reasons and calls tools until it reaches an answer
- Connect any of 140 model providers and 5,162 models through a single adapter contract
- Give agents 50 verified service integrations plus built-in web, fetch, filesystem and shell tools
- Ground answers in your own documents with a one-line RAG pipeline and hybrid vector plus BM25 search
- Persist conversations and embeddings across 17 memory backends, from SQLite to Pinecone
- Render the same chat experience in React, Vue, Svelte, Solid, Angular, React Native or the terminal
- Gate risky tool calls behind human approval, per-tool quotas, sandboxing and prompt-injection detection
When to use AgentsKit.js / When not to
A quick filter to help you decide if AgentsKit.js is the right fit.
When to use AgentsKit.js
- JavaScript and TypeScript developers building their first production agent
- Teams that want to switch model provider, UI framework or memory backend without rewriting their orchestration
- Engineers migrating away from LangChain.js, LangGraph, LlamaIndex, Mastra, OpenAI Assistants or the Vercel AI SDK
- Startups that need a multi-agent runtime, RAG and observability without paying for a hosted platform
- Developers shipping to the edge — Cloudflare Workers, Deno Deploy, Vercel Edge, Bun — where bundle size is a hard constraint
When not to use AgentsKit.js
- Non-technical users looking for a no-code or visual agent builder
- Organisations that require a signed DPA, an SLA or a named legal counterparty before adopting a tool
- Teams that want one library already shipping every third-party integration, where bundle weight does not matter
- Projects that only need streaming chat in React and will never add an autonomous runtime
- Buyers who want a fully prescriptive framework that makes every architectural decision for them
How to use AgentsKit.js
A typical end-to-end flow, from setup to results.
- Install Node.js 20 or newer and create an empty project
- Run npx @agentskit/cli init, or install @agentskit/core and @agentskit/runtime directly
- Write a first agent in a single file with a local adapter and run it — no account, no API key and no network call are required
- Install @agentskit/adapters, replace the local adapter with your provider and set the matching API key in the environment
- Use the stack builder to pick a package manager, framework, provider, memory backend and capabilities, then copy the generated install line and starter code
- Add tools when the agent has to act, taken from the built-ins or the 50 integrations
- Add a memory backend when the agent has to remember across sessions
- Add RAG when answers have to be grounded in your own documents
- Attach observability and a cost guard before sending any production traffic
- Run the evaluation suites and the shipping checklist before changing prompts or models
Pros & Cons
Pros
- Free and MIT-licensed across all twenty-two packages, with no account, no signup and no declared lock-in
- A genuinely small core — around 5.17 KB gzipped, zero dependencies, size budgets enforced in CI
- Six formal contracts make every layer substitutable rather than merely modular
- Unusual breadth for a single foundation: runtime, tools, skills, memory, RAG, UI, observability, evaluation and sandboxing
- Seven UI frameworks, terminal included, behind the same hook
- Documentation written to be read by agents as well as humans, with llms.txt indexes for every property of the ecosystem
- A comparison page that names the cases where the tool is the wrong choice, plus migration guides from six competing frameworks
Cons
- Declared beta: only @agentskit/core is marked stable, the other twenty-one packages are not
- An extremely young project — the domain was registered in April 2026 and the site first archived in August 2026
- No legal entity anywhere on the site: no legal notice, no terms, no privacy policy, no postal address, no email
- No support channel beyond GitHub issues
- No contractual commitment is available: no DPA, no SLA, no subprocessor list
- Translated documentation barely exists — English is complete, Brazilian Portuguese is a landing page, Spanish and Chinese are announced but not mounted
- Strictly developer-facing: everything goes through npm and TypeScript, with nothing no-code
Pricing & Plans
AgentsKit.js is free of charge. All twenty-two @agentskit/* packages are published on npm under the MIT licence, and the site advertises no paid plan, no enterprise edition and no pricing page; there is consequently no entry price point and no currency to report. The only costs are indirect: the tokens billed by whichever model provider the developer connects, and the infrastructure the agent runs on. Both fall outside the publisher's scope, though the library does ship token accounting and a per-run cost ceiling to keep them visible.
- Free — all twenty-two @agentskit/* packages published on npm under the MIT licence
- with no paid tier
Data, GDPR & hosting
A consolidated view of how AgentsKit.js handles your data.
GDPR overview
The site contains no mention of the GDPR whatsoever: no compliance claim, no denial, no Article 27 representative, no data protection officer, no Data Processing Agreement and no subprocessor list. That silence looks structural rather than negligent — the publisher runs no hosted service for this library and is therefore nobody's processor; the controller is whoever deploys the agent. Several documented primitives nonetheless help a compliance programme along: PII redaction before inference and before logging, AES-GCM-256 encrypted chat memory, a hash-chained and HMAC-signed audit log presented as SOC 2 and HIPAA friendly evidence, OIDC and SAML single sign-on, and mandatory sandboxing. None of them amounts to a certification or to a commitment by the publisher.
Who owns the data?
AgentsKit.js publishes no terms and conditions and no privacy policy, so no contractual answer to the ownership question exists on the site. The shape of the product makes that question largely moot: the twenty-two packages are MIT-licensed libraries installed into your own codebase and executed on your own infrastructure, so conversations, documents and embeddings never transit through a vendor service. Ownership and responsibility stay with whoever deploys the agent. In practice data reaches the model provider the developer selects — OpenAI, Anthropic, Google, a local Ollama instance — and is then governed by that provider's terms, not by anything AgentsKit.js says.
Reuse rights
Nothing governs reuse contractually, because AgentsKit.js is distributed as source code rather than as a service. The MIT licence covers the software itself and permits free use, modification, redistribution and commercial exploitation of all twenty-two packages. Whatever an agent produces — transcripts, retrieved chunks, traces, audit logs — is written to storage the developer chooses and controls, so it can be reused without asking anyone. The documented security primitives point the same way: PII redaction strips emails, phone numbers and API keys before they reach the model or the logs, encrypted memory keeps keys off disk in plaintext, and the hash-chained audit log makes tampering detectable. The only external constraint on reuse comes from the model provider selected at runtime.
Data retention & training
Hosting summary
No hosting jurisdiction is published, and none applies in the usual sense: AgentsKit.js is a set of npm packages rather than a hosted service, so the publisher stores nothing on a user's behalf. Wherever the agent runs is where the data lives. The documentation covers Node, Bun, Cloudflare Workers, Deno Deploy and Vercel Edge as deployment targets, plus a browser-only pattern in which the model itself executes locally through WebGPU and, in the project's own words, no inference data ever leaves the device. Persistence is equally the developer's choice, spread across seventeen supported backends including SQLite, Redis, Postgres with pgvector, Turso, Upstash, Pinecone, Qdrant, Chroma, Weaviate, Milvus, MongoDB Atlas and Supabase. Each carries its own jurisdiction, and that jurisdiction is the one that counts. The only hosting fact attributable to the project itself is its documentation site, served from a United States address on Amazon infrastructure — a marketing site, not a data platform. Anyone with residency obligations therefore sets them at the storage and provider layer: the library imposes nothing and guarantees nothing.
Things to keep in mind
Risks and trade-offs to weigh before adopting AgentsKit.js.
- With no legal entity, no terms and no privacy policy, there is nobody to notify and nothing to invoke if something goes wrong
- Beta status on twenty-one of the twenty-two packages means contracts outside the core can still change under a running project
- Autonomous agents act on real systems: a tool call that writes to production is only as safe as the quotas, sandbox and approval gates you actually switch on
- Token costs are billed by the model provider and stay uncapped unless the developer enables the cost guard
- The documentation speaks of honest isolation claims for the sandbox — real isolation depends on the backend chosen, and untrusted code remains untrusted code
- Delegating reasoning to an agent loop erodes the habit of reading the intermediate steps; traces exist precisely because the final answer alone should not be trusted
- A project this young can slow down or stop, and an MIT licence protects the code you already have rather than the roadmap you were counting on
Setup & Integrations
Technical difficulty
Moderate, and honest about it. Node.js 20 and working TypeScript are the real prerequisites; there is no visual builder and no hosted console. The start is genuinely quick — one install command, a first agent that runs from a single file without an account or an API key, and a streaming chat claimed in under ten lines. A stack builder generates the install line and starter code, a five-step browser tutorial needs no keys, and agentskit doctor diagnoses providers and versions. Production is the harder half: observability, evaluations, guardrails and sandboxing are separate, deliberate steps.
Deployment
Integrations
Supported languages
Behind AgentsKit.js
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement AgentsKit.js.
Frequently asked questions
How much does AgentsKit.js cost?
Do I need an account or an API key to try it?
Which models and providers can I connect?
Which front-end frameworks are supported?
What do I need to know before starting?
Is it ready for production?
What happens to my data?
Is there any support?
Can I migrate from another framework?
In which languages is the documentation available?
Should you pick AgentsKit.js?
AgentsKit.js makes an unusually precise bet: that the problem with JavaScript agent development is not missing features but accumulated glue. Its answer is six formal contracts on a core of roughly 5.17 KB, twenty-two packages installed one at a time, and a licence that asks nothing in return. The breadth is real — runtime, tools, skills, memory, RAG, seven UI bindings, observability, evaluation, sandboxing — and so is the substitutability, which is the part most frameworks promise and few deliver.
The candour helps. A comparison page that names three situations where the tool is the wrong choice, migration guides written from six competing frameworks, and a first agent that runs without an account or an API key all suggest a project confident enough to be tested rather than believed.
What holds it back is age and absence. The domain was registered in April 2026 and the site first archived in August 2026; twenty-one of the twenty-two packages carry a beta label. More consequential for anyone evaluating it professionally, there is no publisher to speak of — no legal notice, no terms of service, no privacy policy, no postal address, no email, no support channel outside GitHub issues. Nothing can be signed and nobody can be notified.
That combination points to a clear reading. For a team that already writes TypeScript, wants to keep control of its stack, and treats a framework as source it owns rather than a vendor it depends on, AgentsKit.js is an inexpensive and technically serious option. For an organisation that needs a Data Processing Agreement, a service level commitment or a named counterparty before adopting anything, it is not yet a candidate — and the project makes no claim otherwise.
- Choosing a selection results in a full page refresh.
- Opens in a new window.