AgentsKit - The most complete ecosystem to create AI agents

👋 Emerson here, solo maintainer. Short story. Earlier this year I tried to build a simple AI chat in JavaScript — streaming, 2 tools, memory across sessions. Not curing cancer, just a chat. I ended up gluing 5 libs together: • Vercel AI SDK for chat (no runtime) • LangChain.js for runtime (200MB, leaky abstractions) • assistant-ui for components (53 of them) • MCP for tool interop (no UI) • Pinecone bolted on 3 days in, couldn't debug it. 5 days in, deleted everything. 6 months later → AgentsKit. What's different: → 10KB zero-dep core, bundle gates in CI → 6 formal contracts (Adapter, Tool, Memory, Retriever, Skill, Runtime) — published as ADRs, not magical abstractions → Same `useChat()` works in React AND terminal (Ink) → Swap providers in 1 line — 8+ adapters, one contract → Built-in: tools, memory, RAG, sandbox, observability, eval → `npx agentskit init` scaffolds a working agent in 30 seconds Shipping now (v1.0): • 14 packages on npm, 538 tests • Full docs at agentskit.io • 4 starter templates (react, ink, runtime, multi-agent) Shipping Phase 2 this month: • Deterministic replay for LLM calls (record seed + responses, reproduce bit-for-bit) • Prompt diff tool (git blame for prompts) • Time-travel debugger (rewind traces, alter tool outputs, re-run) • Token budget compiler Honest comparison: • vs Vercel AI SDK — their chat UI rules, their runtime is missing • vs LangChain.js — 200x smaller, no leaky abstractions • vs Mastra — solid runtime, we cover more surface (browser + terminal + CLI) • vs MCP — we consume MCP servers; MCP is a protocol, AgentsKit is the kit Free, MIT, built in the open. No cloud signup. No telemetry you can't turn off. Links: 🌐 https://www.agentskit.io 💻 https://github.com/EmersonBraun/... 💬 https://discord.gg/zx6z2p4jVb (Office Hours every Friday) Brutal feedback welcome — especially where the pitch overclaims or where position vs the ecosystem is unclear. Will be here all day 🙏