EnterpriseAgentFramework (ReachAI)

What it is – ReachAI is an open‑source platform that lets Java‑based enterprise applications (OA, ERP, CRM, eHR, procurement, ticketing, contracts, etc.) embed AI assistants directly into their existing user interfaces. The AI can read the current page, query data, fill forms and trigger business actions while respecting the user’s identity, permissions and the original workflow.

---

Core ideas

Idea	How ReachAI implements it
AI‑first retro‑fit	Scans a running Java system, builds a map of pages, routes, components and backend APIs, then suggests up‑to‑three concrete automation candidates.
AI‑Coding	Turns each candidate into a structured engineering task (scope, code context, constraints, acceptance criteria) and hands it to LLM‑based code generators (Codex, Claude‑Code, Cursor, etc.). The generated code is committed, tested and verified before the task is closed.
Capability catalog	Java methods annotated with @ReachCapability are automatically registered as reusable, versioned business capabilities (read/write, required roles, side‑effect level).
Controlled execution	An Agent interprets a natural‑language request, selects a pre‑approved Workflow (GraphSpec) from a whitelist, and runs it through a runtime that records every step (Trace). Critical write actions require explicit user confirmation.
Explicit page bridge	Instead of screen‑scraping or mouse‑simulation, the front‑end registers readable, filterable and writable actions for each page. The AI calls these actions via a typed contract.
Governance & observability	Operations dashboard shows agents, runs, users, token usage, risk items and business impact. RunOps, EvalOps and Automation modules provide automated testing, version‑controlled datasets and replay of any execution.
Open ecosystem	Capabilities, Tools and Workflows can be published to an internal API market, MCP hub or A2A hub, enabling other services or remote agents to call them under the same ACL and audit framework.

---

Main components (five services)

Service	Port	Role
reachai-model-service	18601	Model gateway – chat, embedding, rerank, model instance management
reachai-knowledge-service	18602	Knowledge base – documents, vector store (Milvus), RAG indexing (/ai path)
reachai-control-service	18603	Public API / BFF – project & page management, auth, embed endpoints
reachai-runtime-service	18604	Runtime host – Agent supervisor, Workflow executor, Trace & RunOps
reachai-capability-service	18605	Capability catalog – SDK registration, snapshots, diffs, review

---

Typical workflow

1. Connect a Java business system (SDK starter or passive scan).
2. Scan pages, routes, components and backend APIs → generate a system capability map.
3. Recommend up to three automation ideas (read‑only analysis, fact/inference tagging).
4. AI‑Coding creates a task, LLM writes/modifies code in the real repo, runs tests, and returns evidence.
5. Publish the new @ReachCapability and/or a Workflow (GraphSpec) that the Agent can invoke.
6. Run – users type natural language in the embedded chat; the Agent selects a whitelisted Workflow, executes it, asks for confirmation on writes, and logs the full trace.
7. Govern – Ops dashboard aggregates usage, risk, token consumption and version performance; Automation can schedule periodic runs, EvalOps can run A/B experiments on new versions.

---

Who benefits?

Role	Benefit
Developers	Automated scanning & AI‑generated code reduces manual integration effort; SDK gives a typed way to expose existing services as AI‑callable capabilities.
Business users	No UI change – they stay in their familiar OA/CRM page and ask the AI to “show my pending tasks” or “create a purchase request”.
Managers / Ops	Real‑time view of AI adoption, risk items, token usage and business impact; ability to plan the next round of automation based on actual usage data.

---

Technology stack

• Backend: Java 17, Spring Boot 3.4.5, Spring AI, Spring AI Alibaba extensions
• Agent/Workflow engine: AgentScope 2.0, LangGraph4j 1.8, GraphSpec DSL
• SDK / Page bridge: JDK 8‑compatible annotations, Spring Boot 2 starter, TypeScript embed SDK
• Frontend: Vue 3.5 + TypeScript, Element Plus, Pinia, Vue‑Flow, Vite 6
• Infra: MySQL 8, Redis 7, Milvus 2.4 (vector store), optional Docling for document ingestion

---

Maturity & License

• The repository is actively maintained (README mentions fast‑iteration, detailed service boundaries and local‑dev scripts).
• All core features described (scanning, AI‑coding, capability catalog, workflow studio, runtime, governance) are implemented, though some workflow nodes are marked beta.
• Licensed under the MIT License, allowing commercial use and modification.

---

Getting started (quick summary)

1. Install JDK 17, Maven, Node ≥18, Docker‑Compose.
2. cp deploy/.env.example deploy/.env → edit secrets.
3. docker compose -f deploy/docker-compose.infra.yml up -d and run the provided SQL init script.
4. Build backend: mvn clean install -DskipTests.
5. Launch the five services (IntelliJ run configuration or mvn spring-boot:run in the order shown).
6. Front‑end: cd ai-admin-front && npm ci && npm run dev.
7. Log in with the default dev credentials (admin / admin123).

---

Bottom line

ReachAI (EnterpriseAgentFramework) is a full‑stack, production‑oriented framework for turning existing Java enterprise applications into AI‑augmented assistants. It combines automatic system discovery, LLM‑driven code generation, a typed capability registry, safe workflow execution, and extensive governance tooling—exactly the kind of end‑to‑end solution needed for enterprise AI adoption.