AgentsKitVerified
Open-source, MIT-licensed JS/TS toolkit for building AI agents with a 10KB zero-dependency core and 50 built-in integrations.
- Category
- Agent frameworks
- Pricing model
- free
- Free tier
- Yes
- Deployment
- Self-host
- Interface
- Library
- Open source
- Yes
- Public API
- Yes
- Model / LLM
- Provider-agnostic: Anthropic, Op…
Facts re-verified
On this page
Our verdict
AgentsKit is a lean, code-first way to assemble AI agents in JavaScript or TypeScript without pulling in a heavyweight framework or a hosted platform.
The zero-dependency, sub-10KB core and the Registry's copy-not-install model (you get owned TypeScript source, not a hidden runtime) are genuinely developer-friendly design choices, and the Doc Bridge MCP server addresses a real problem: coding agents editing against stale docs.
The catch is early-stage signal: 25 GitHub stars, a core still at v1.0.0 while the rest of the ecosystem grows, and public feature counts that don't quite match between the vendor's own site and its README. Worth trying for a side project or an internal tool; a team betting production infrastructure on it should read the source directly rather than take the numbers at face value.
What is AgentsKit?
AgentsKit is an open-source, MIT-licensed toolkit for building AI agents in JavaScript and TypeScript, built around a core package under 10KB with zero runtime dependencies. It ships as 22 npm packages covering 25 native LLM adapters, 50 built-in integrations (Slack, Notion, Stripe, GitHub, Shopify among them), 17 memory backends and 7 UI framework bindings for React, Ink and headless setups.
A companion AgentsKit Registry lets developers pull ready-made, typed agent factories straight into their codebase via a CLI, and a separate Doc Bridge MCP server plugs into Cursor, Claude Code and Copilot to keep AI coding agents in sync with project docs. It serves JS/TS developers who want to assemble or extend agents without adopting a hosted platform.
What does AgentsKit do?
Developers install AgentsKit by running npm install @agentskit/core @agentskit/runtime tsx, or scaffold a new project with npx @agentskit/cli init. The core requires Node 22+ and TypeScript 5.8+, and it swaps between LLM providers, Anthropic, OpenAI, Gemini, Ollama, DeepSeek and Grok among them, by changing one line of adapter configuration. Because the core is under 10KB with zero dependencies, it can run without any LLM or API key at all for local development, and no telemetry or project data is sent back to AgentsKit.
Beyond the base runtime, the AgentsKit Registry offers ready-to-use agents that are copied, not installed as a dependency: running npx agentskit add <id> downloads a manifest from /r/<id>.json and writes the agent's source directly into ./agents/<id>/ in the project. Each registry entry is a provider-agnostic TypeScript factory with a typed contract, tests and evaluation coverage, so there is no shared registry runtime and no silent version drift; updates only happen when a developer explicitly re-pulls.
A separate component, Doc Bridge (installed via npm i -D @agentskit/doc-bridge, CLI ak-docs), runs as an MCP server exposing eight read-only tools such as handoff.resolve and doc.search so that Cursor, Claude Code, Copilot and other MCP-compatible agents can check project documentation before editing code, with a GitHub Action available to catch stale docs in CI.
How AgentsKit works
- Run npx @agentskit/cli init, or npm install @agentskit/core @agentskit/runtime tsx, to scaffold a project on the zero-dependency core.
- Pick an LLM adapter from the 25 supported providers and wire it in with one line, or run the core with no API key for local testing.
- Pull a ready-made agent with npx agentskit add <id>, which downloads its manifest and writes typed source into ./agents/<id>/ for you to own and edit.
- Install Doc Bridge (npm i -D @agentskit/doc-bridge) if using Cursor, Claude Code or Copilot, so those agents check handoff.resolve and doc.search before editing project files.
Key features
- Zero-dependency core
- Core package stays under 10KB gzipped with zero runtime dependencies, enforced by CI on every release, and can run with no LLM or API key at all for local testing.
- 25 native LLM adapters
- Swap between providers including Anthropic, OpenAI, Gemini, Ollama, DeepSeek and Grok by changing one line of adapter configuration.
- AgentsKit Registry
- Copy typed, tested agent factories directly into your own codebase with npx agentskit add <id>; there is no shared registry runtime, so nothing updates silently.
- Doc Bridge MCP server
- Exposes eight read-only tools (handoff.resolve, doc.search, doc.get, gate.status among them) so Cursor, Claude Code and Copilot check project docs before editing.
- 50 built-in integrations
- Ships with connectors for Slack, Notion, Stripe, GitHub and Shopify among 50 total, spread across 22 published npm packages.
- Multi-surface UI bindings
- Supports React, Ink (terminal) and headless usage across 7 supported UI frameworks, plus 17 memory backends for state.
What are AgentsKit's use cases?
- Scaffolding a new agent project
- A developer runs npx @agentskit/cli init to get a working TypeScript agent skeleton wired to a chosen LLM provider in minutes.
- Reusing a pre-built agent instead of writing one
- A team runs npx agentskit add <id> to copy a tested, typed agent factory into their repo, then edits the owned source directly instead of building from scratch.
- Keeping coding agents doc-aware
- A team installs Doc Bridge so Cursor or Claude Code calls handoff.resolve before editing, reducing changes made against stale project documentation.
- Shipping a terminal-native agent interface
- A developer uses the Ink UI binding to build a CLI agent experience instead of standing up a web dashboard.
Who is AgentsKit for?
- JavaScript/TypeScript developers building custom AI agents who want a lightweight, dependency-free core instead of a heavyweight framework.
- Teams using Cursor, Claude Code or GitHub Copilot who want those tools to verify project documentation before making edits.
- Developers who want agent code they can copy, own and modify outright, rather than call through a hosted API.
Not forNon-developers or teams wanting a no-code, hosted agent builder: AgentsKit requires Node 22+, TypeScript, and comfort reading generated source code. The open-source core has no built-in managed dashboard or SLA of its own — teams that want one can use AKOS, a separate paid enterprise control plane for governance, audit and cost controls, but it is optional and does not gate the open-source stack.
What does AgentsKit integrate with?
- Slack
- Notion
- Stripe
- GitHub
- Shopify
- Cursor
- Claude Code
- GitHub Copilot
- ClawHub
- Pi
- Anthropic
- OpenAI
- Gemini
- Ollama
- DeepSeek
- Grok
Why use AgentsKit?
- Core package stays under 10KB gzipped with zero runtime dependencies, checked by a CI-enforced budget on every release.
- Registry agents arrive as typed TypeScript factories with tests and evaluation coverage that you own outright, not calls through a hidden hosted runtime.
- Works entirely offline for local development: no LLM, API key, account or telemetry required to get started.
- Provider swapping across Anthropic, OpenAI, Gemini, Ollama, DeepSeek and Grok is a one-line adapter change rather than a rewrite.
What are AgentsKit's pros and cons?
What's great
- Core package is under 10KB gzipped with zero runtime dependencies and a CI-enforced size budget, adding negligible weight to a project.
- Registry agents are copied as owned, typed TypeScript source with tests and eval coverage rather than pulled through a hidden runtime, avoiding lock-in.
- Works without any LLM, API key or account for local development, and sends no telemetry back to AgentsKit.
- Provider swapping across Anthropic, OpenAI, Gemini, Ollama, DeepSeek and Grok is a one-line change rather than a rewrite.
Watch-outs
- Public feature counts are inconsistent across the vendor's own pages: the main site cites 184 providers and 6,297 models, while the GitHub README's evidence ledger lists 140 providers and 5,000+ models.
- Adoption signal is thin: the GitHub repository shows only 25 stars and 1 fork at review time.
- Requires Node 22+ and TypeScript 5.8+, a recent toolchain that can exclude teams running older stacks.
- The Doc Bridge Claude Desktop MCP Bundle currently only packages for macOS.
- AKOS, the vendor's optional paid managed control plane for governance, audit and cost controls, is still in early access with no published pricing — interested teams must book a demo to learn cost.
- Details here come from the vendor's own site, its GitHub repository, and the mcpservers.org directory listing; independent editorial reviews are not available yet.
- The vendor describes @agentskit/core as v1.0.0 while stating the rest of the ecosystem (adapters, integrations, memory backends) is still growing, so maturity likely varies package by package.
Frequently asked questions
Is AgentsKit free to use?
What runtime does it require?
Does pulling an agent from the Registry create a dependency on AgentsKit?
Which LLM providers does it support?
What is Doc Bridge and does it work everywhere?
Anything to add?
This page is researched and re-verified by our editors — but the people who actually use AgentsKit know things we cannot see from the outside.
Report an issueSomething wrong, outdated or a dead link? Tell us and we re-verify.
Write a reviewUsed AgentsKit? Your experience helps the next buyer decide.
Get a quote from AgentsKitTell us what you need and we make the introduction.
Get weekly Agent frameworks updates
New agents and fresh verdicts. One short email, every week. One-click unsubscribe.