AI Agent Builder — Build with natural language

Build AI agents visually. Or just describe them.

Agent Studio is a no-code builder on top of agentsdk-js — drag-drop on the canvas, or just tell the AI assistant what you want and watch it build the workflow for you. Multi-provider, multi-agent, deploy anywhere.

Works withGLMOpenAIAnthropicDeepSeekGeminiBYOK
39
Node types
Trigger · Agent · Media · Meta · Web · …
5+
Providers
GLM · OpenAI · Anthropic · DeepSeek · Gemini
20
Templates
Research · Marketing · Support · …
MIT
SDK license
Open source forever
Features

Everything you need to ship agents

Six capabilities that come standard — no extra dependencies, no extra config. Built into the same runtime that powers the no-code builder.

AI assistant builds it for you

Just describe what you want — "build me a marketing agent that posts to Instagram" — and the AI assistant generates the entire workflow on the canvas. Edit, refine, or run it instantly.

Trigger
Agent
Output

AI agents as first-class nodes

Agents live on the canvas like any other node — wire them up, inspect, run, and reuse. No special casing, no boilerplate.

Multi-agent orchestration

One agent calls another as a tool — recursive handoff built into the runtime.

No-code builder

Drag-and-drop React Flow canvas with 39 node types — design graphs visually, inspect each node, run live, and export to JSON or push to GitHub.

HTTP Trigger
Router Agent
Cost + Output

Multi-provider BYOK

GLM, OpenAI, Anthropic, DeepSeek and Gemini — bring your own key, swap at will.

Cost tracking + deploy anywhere

Every run returns token usage and USD cost. Export as a standalone Node.js project, push to GitHub, or deploy to Railway / Vercel / Docker in one click.

Architecture

5-layer architecture

Each layer is independently usable. Build on whichever layer fits your problem — from a one-liner generateText call to a full multi-agent graph.

Platform Layer

Phase 5 · planned

Auth, billing, deploy, scheduled triggers — the hosted product surface.

No-Code Builder UI

Phase 4 · done

React Flow canvas, node palette, inspector, run panel — what you see today.

Node Graph Engine

Phase 3 · done

JSON schema, validation, topological execution engine with cost rollup.

Agent Runtime

Phase 2 · done

Step loop, tool calls, agent-as-tool handoff, memory and streaming.

Core SDK

Phase 1 · done

Provider adapters, generate/stream/object functions, cost + error model.

Quick start

Use it in code

Three lines to register a provider, one call to generate. Costs and usage come back automatically — no observability vendor required.

quick-start.ts
import { generateText, createGlmProvider, registerProvider } from 'agentsdk-js' registerProvider(createGlmProvider()) const result = await generateText({  model: 'glm-4.6',  messages: [{ role: 'user', content: 'What is 2+2?' }],}) console.log(result.text)        // "4"console.log(result.cost.total)  // 0.000036
or

Or use the no-code builder

Skip the boilerplate. Drag, drop, wire, run — export to JSON when you are done. Same runtime, same graphs, zero config.

Packages

One unified package — three layers

Install a single npm package, agentsdk-js, and get the core SDK, agent runtime, and graph engine. MIT licensed, fully typed, zero runtime deps.

MIT
Layer 1 · Core
agentsdk-js

Core SDK — generateText / streamText / generateObject, provider registry, cost tracking, typed errors, built-in GLM + OpenAI + Anthropic providers.

MIT
Layer 2 · Runtime
agentsdk-js

Agent runtime — step loop, tool calls, agent-as-tool handoff, in-memory + custom memory strategies.

MIT
Layer 3 · Graph
agentsdk-js

Graph engine — JSON graph schema, validation, topological execution with {{ref}} variable resolution. The no-code runtime.

$bun addagentsdk-js

Start building agents today

Open the studio, drop a few nodes on the canvas — or just ask the AI assistant to build it for you — and run your first agent in under a minute.

Open source SDKMIT licensed39 node typesAI assistant