Build AI agents on a canvas.
Ship them as real code.
AgentsKit turns agents, tools, memory, MCP servers and multi-agent logic into blocks you connect visually — then tests them against real models and exports a clean Node.js project.
No sign-up. Runs in your browser. Free.
Not another chatbot
A development environment for agents
Everything you'd wire up by hand in LangGraph — represented as structured, validated graph data you can see and edit.
Drag, drop, connect
Pull agents, models, tools and memory from a searchable library. Connect them left-to-right for execution order and bottom-up for resources.
108 real components
Agents, sub-agents, DeepAgents, routers, loops, human approval, RAG pipelines, MCP servers, skills and plugins — each with a focused config panel.
Test with real models
Run the workflow against OpenAI, Anthropic or Gemini and watch it execute node by node, with tool calls, outputs and errors. Nothing is simulated.
Export TypeScript or JavaScript
Get a strictly-typed TypeScript (or plain JS) project — models, tools, agents and a LangGraph StateGraph that mirrors your canvas — ready for npm start.
Validation built in
Missing models, dangling branches, duplicate tool names, approvals without a checkpointer — problems are flagged before you run or export.
Local-first, no sign-up
Start building immediately. Projects and history live in your browser; back them up as JSON whenever you like.
Built on the official SDKs
The LangChain ecosystem, as building blocks
Every component maps to a documented API. Integrations that can't run in a hosted test (local files, databases, stdio MCP) are labelled clearly and still export as working code.
LangChain.js
createAgent, middleware, 17 chat model providers via initChatModel, and 30 prebuilt tools.
LangGraph.js
StateGraph, conditional edges, checkpointers, stores and interrupts.
Deep Agents
createDeepAgent with planning, a virtual filesystem, sub-agents and skills.
MCP
Connect stdio or Streamable HTTP servers with @langchain/mcp-adapters.
How AgentsKit works
- 1
Start from a template or a blank canvas
No account. Your project is created instantly and autosaved locally.
- 2
Compose your workflow
Drop components, connect them, and configure each one in plain forms — no JSON.
- 3
Validate and test
Fix highlighted problems, add an API key, and run it for real in the test panel.
- 4
Export and ship
Download a runnable Node.js project with dependencies, .env example and README.
Node.js code export
Your canvas, as code you'd write yourself
- Separate files for models, tools, MCP, memory, agents and the graph
- Routing, loops, parallel branches and human approval as LangGraph edges
- package.json with pinned dependencies and a .env.example
- README with setup and run instructions — secrets never embedded
1export const graph = new StateGraph(State)2 .addNode("input", inputNode)3 .addNode("intent_router", intentRouterNode)4 .addNode("billing_agent", billingAgentNode)5 .addNode("technical_agent", technicalAgentNode)6 .addNode("output", outputNode)7 .addEdge(START, "input")8 .addEdge("input", "intent_router")9 .addConditionalEdges("intent_router", (state) => intentRouterBranches[state.route] ?? [END], ["billing_agent", "technical_agent", END])10 .addEdge("billing_agent", "output")11 .addEdge("technical_agent", "output")12 .addEdge("output", END)13 .compile({ checkpointer: memory.checkpointer });Example workflows
Load any of these as an editable project.
Extract & Route
Pull structured fields out of a message into variables, branch on them, and write a reply that uses them.
Simple AI Assistant
A single agent with a model and conversation memory. The best place to start.
Customer Support Agent
An LLM router sends billing and technical questions to specialised agents with their own tools.
Research Agent
A ReAct agent that searches the web, reasons over results and writes a sourced answer.
Build your first agent in minutes
Open the dashboard, pick a template, and export working code — no account required.
Start building