RocketChain

Visual editor for agentic flows that compiles to Spring Boot. Think LabVIEW for AI agents: drag nodes onto a canvas, wire them with exec and data edges, click Generate — out comes a production-grade Java backend powered by agentic-helper.

🌐 Live app: rocketchain.vercel.app

There is no runtime server — everything runs in your browser via the File System Access API. Your project lives on disk; RocketChain just edits it.

---

Why

Building agentic backends in Java means stitching together LLM calls, control flow, retries, async coordination, validation, and Spring Boot scaffolding. RocketChain pulls all of that into a typed visual graph: each node is a small typed unit (Agent, IfElse, ForEach, SubflowCall, ...), edges are typed too (orange = control, blue = data), and the codegen lowers the whole thing to a @Service plus a typed trigger wrapper (@RestController / @Scheduled / @EventListener).

It's also designed to be agent-friendly: every .rocketflow is plain JSON, validated by a JSON Schema. The included AGENTS.md translation guide lets an AI coding assistant author flows by hand, and the codegen CLI gives a fast feedback loop (mvn compile in ~5s).

---

Quick start

1. Open the live app at rocketchain.vercel.app, or clone & run locally:

   git clone https://github.com/Yann-Favin-Leveque/rocketchain.git
   cd rocketchain
   npm install
   npm run dev

2. Create a project — New project on the home page walks you through identity, database, security, and LLM providers. The wizard scaffolds the directory, the .rocketchain manifest, and an .env skeleton.
3. Configure LLM_INSTANCES — open Project Settings → LLM Providers and fill in at least one provider (OpenAI, Anthropic, Mistral, Grok, DeepSeek, Gemini, or Azure flavors of each).
4. Build a flow — create an agent JSON, drop it into a .rocketflow, wire Start → Agent → End, click Generate. The output is a mvn clean compile-ready Spring Boot project.

---

Core concepts

A RocketChain project is a directory on disk containing:

• *.rocketflow — a typed graph (ReactFlow JSON), type: "main" or type: "sub"
• prompts/agents/*.json — one LLM persona per file (model, instructions, tools, ...)
• prompts/blocks/*.json — reusable prompt fragments inlined via <<<block::name>>>
• prompts/skeletons/*.json — prompt templates with editable slots <<{name}>>
• prompts/tools/*.json — function-call tool definitions for autonomous agents
• src/main/java/<base>/... — generated Spring Boot code (services, controllers, tools)

The visual editor manages all of this; nothing forces you to use the UI — files are plain JSON / Java and a JSON Schema (schema/rocketflow.schema.json) validates them.

---

Visual editor

ReactFlow canvas with two edge kinds:

• Exec edges (orange) — control flow (which node runs next)
• Data edges (blue) — typed values between handles

Right-click for the node palette. Each category has an accent color:

Category	Examples
Agent	AgentNode — fires an LLM call, optional autonomous + tools
Data	ValueNode, VariableNode, GlobalVariableNode
Operator	math, comparison, casts, list/string ops, boolean
Control flow	If/Else, For, While, Switch, Try/Catch
Async	CompletableNode, forCompletable, listCompletable, AllOf
Flow	Start, End, Input, Output, SubflowNode, EventEmitter

InputNode / OutputNode only appear in sub-flows; Start / End only in main flows.

---

Prompt DSL

Agents and blocks support a small inline DSL:

Tag	Effect
<<<block::name>>>	Inline a reusable block at this position
<<<input_format>>>	Injected at runtime: the agent's input JSON schema
<<<output_format>>>	Injected at runtime: the agent's output JSON schema
<<<var::X>>>	Dynamic prompt variable, wired at runtime via an input handle on the AgentNode
<<<globalvar::X>>>	Resolves at codegen time from GlobalConstants.java
<<{slot}>>	Skeleton-editable region — only these are mutable when an agent uses the skeleton

The prompt editor has an Expand toggle to preview the fully-resolved prompt with blocks, skeletons and format placeholders all merged.

---

LLM providers

Configure in Project Settings → LLM Providers. Supported:

openai · anthropic · mistral · grok · deepseek · gemini · azure-openai · azure-anthropic · azure-mistral · azure-grok · custom (data-driven CustomProviderSpec)

Per-instance fields: id, provider, url, key, models (CSV), enabled, optional apiVersion (Azure) and rateLimits ({ "*": 3 } or per-model { "gpt-5": 40 }).

Keys can be inlined as literals or referenced via ${VAR} placeholders resolved at boot by spring-dotenv.

---

Code generation

Click Generate. The pipeline:

1. Validation — blocks generation on errors (e.g. agent with model="none", sub-flow without Input/Output, undefined tool reference). Warnings are reported but don't block.
2. Compile — each main .rocketflow lowers to a @Service + a typed trigger wrapper (@RestController / @Scheduled / @EventListener). Sub-flows inline as private helper methods (multi-output → static records).
3. Write — Java files land under src/main/java/<base>/{service,controller,tools}/. Tab-driven Java families (Entity, Repository, Routine, Event) are merged via //{Edit here:} markers, preserving user edits across regen.
4. Stack — Spring Boot 3.4 · Java 17 · agentic-helper 1.24+ · Hibernate 6 · Spring Security 6 · JJWT 0.12.

After generation, mvn clean compile should pass.

---

Authoring .rocketflow by hand

.rocketflow files are plain JSON validated by schema/rocketflow.schema.json (auto-loaded in VS Code via .vscode/settings.json). The full translation guide for AI coding assistants — methodology, node catalog, agent/tool shapes, common pitfalls — lives in AGENTS.md.

A CLI is provided for fast iteration without the UI:

# Stage-by-stage debug dump (parser → walker → AST → emitted Java)
npx tsx tools/codegen-cli.ts debug path/to/flow.rocketflow

# Full project build + mvn compile
npx tsx tools/codegen-cli.ts smoke path/to/flow.rocketflow \
  --subflows=path/to/subflows-dir \
  --agents=path/to/agents-dir \
  --tools=path/to/tools-dir \
  --out=tmp/out --package=com.example

---

Sub-flows

Reusable typed graphs callable from main flows via SubflowNode. Rules:

• Exactly one InputNode and one OutputNode per sub-flow.
• No Start/End in sub-flows; no Input/Output in main flows.
• InputNode's named/typed parameters → sub-flow signature; OutputNode's → return type.

Multi-output sub-flows lower to a static record:

SubflowResult_processOrder result = _sub_processOrder(amount);
Double total = result.total();
String label = result.label();

The compiler resolves transitively-reachable sub-flows automatically.

---

Tech stack

Frontend — React 18.3, Vite 6, TypeScript 5, ReactFlow 11, Tailwind, shadcn/ui, Geist design system.

Codegen — pure TypeScript pipeline (parser → walker → AST → emitters). Browser-side via the File System Access API; no backend.

Generated runtime — Spring Boot 3.4 + agentic-helper (multi-provider Java AI library, supports prompt caching, tool-use loops, structured outputs, custom providers).

---

Project status

v1.0 — public beta. The visual editor and codegen pipeline are functional end-to-end. Areas of active iteration:

• Marker-preservation coverage across more file families
• Performance / code-splitting on the frontend bundle
• Additional node types and sub-flow ergonomics

Issues and feature requests welcome on the GitHub repo.

---

License

MIT — see LICENSE. Built by Yann Favin-Levêque.