AgentsKit

AgentsKit is an open-source JavaScript AI agent framework that gives developers a unified ecosystem for building, running, and scaling AI agents without assembling multiple disconnected libraries. Built around a 10KB zero-dependency core, it ships 19 focused npm packages covering chat UI, runtime, tools, memory, RAG, observability, and production guardrails — install only what a project needs. The architecture is built around six formal contracts, meaning every adapter, skill, memory backend, and retriever is substitutable: switching from OpenAI to Anthropic or from a React chat UI to a terminal interface requires changing one line, not rewriting prompt plumbing. AgentsKit supports 20-plus LLM providers including OpenAI, Anthropic, Gemini, Grok, Ollama, and DeepSeek through the same adapter interface. It runs identically in browser, Node.js, Deno, and Bun, and provides framework bindings for React, Vue, Svelte, Solid, React Native, Angular, and Ink. The framework is agent-first rather than chat-first: ReAct loops, reflection, planning, and multi-agent delegation are built into the runtime. Tools — covering browser, filesystem, search, email, and code execution — enforce a strict contract that supports parallel calls and human approval steps. Observability is built in via LangSmith, OpenTelemetry, and console tracing covering every LLM call, tool invocation, and memory operation. A CLI scaffolding command (npx @agentskit/cli init) generates a working agent from four templates in under a minute. The project is MIT-licensed and reports roughly 11,000 monthly npm downloads across its packages.

Building JavaScript AI agents with multi-LLM provider support, creating customer support agents with memory and CRM tool integration, developing research agents with web search and summarisation workflows, building code automation agents with filesystem and runtime tools, adding RAG-powered internal copilots to existing JS applications, running multi-agent delegation and planning workflows, deploying agents across browser and server environments from a single codebase, prototyping agents quickly with CLI scaffolding templates, integrating LangSmith or OpenTelemetry observability into agent pipelines