EvolvIQ (智衍) · AI-Native Industrial Agent Platform

The world's first open-source, AI-native industrial agent platform — 25 pre-built agents spanning L2 (shop-floor protocols) to L4 (enterprise decision intelligence), designed for electronics manufacturing and semiconductors.

🏷 Latest release: v20.5 — Production Data Layer (P1) + Multi-Tenant & Live Graph (P2) (2026-08-05) · 25 Agents · Memory (P0) + Self-Learning (P1) + Self-Evolution (P2) + Production Data Layer closed loop

📖 User Guide: docs/GUIDE.md · 中文指南

✨ Features

• 25 Industrial Agents: Pre-built autonomous agents for supply chain, R&D, manufacturing, quality, and enterprise decision-making — spanning shop-floor protocols to enterprise decision intelligence
• 65 MCP Tools: Standardized tool federation via the Model Context Protocol (HTTP + stdio dual transport)
• 4 Industrial Protocol Gateways: Modbus, MQTT, OPC-UA, IPC-CFX — real or simulated mode
• Multi-Agent Orchestration: 8 preset collaboration templates (NPI / OEE / Quality / Energy / ECO ...) — automatic goal decomposition, parallel agent execution, and cross-agent insight aggregation
• Cross-Agent Knowledge Graph: Neo4j-backed semantic network with automatic in-memory fallback
• Experience Memory & Recall (P0): Every agent execution + multi-agent orchestration writes back cross-agent insights (Insight nodes); agents recall relevant history before reasoning via BaseAgent.recall(goal) / /kg/recall — the memory loop is closed
• Persistent Effects & Audit: Metrics (autonomy rate, time saved) and audit logs survive restart via SQLite fallback — effect-driven tuning builds on real history, not a blank slate
• Authorization Engine: Per-agent confidence thresholds, daily autonomy limits, and approval boundaries — real-time AI behavior guardrails
• Graceful Degradation: Every external dependency (PostgreSQL, Neo4j, OPC-UA Server, AMQP Broker) automatically degrades to local alternatives — never blocks startup or execution
• Multi-Tenant: Row-level tenant_id isolation with API-Key authentication (X-Tenant-Key header)
• Effect-Driven Strategy Tuning: Live knob adjustment (confidence thresholds, daily limits) with audit trail
• Self-Learning Loop (P1): Human approve/reject in the Intervention Center auto-feeds an agent's preference/forbidden memory; the strategy tuner auto-adjusts guardrails (with one-click rollback) — the system learns from experience, it does not just execute
• Self-Evolution Loop (P2): LLM replays human-rejected cases to propose a revised agent system prompt — versioned, human-approved (never auto-applied), with hot-swap + one-click rollback; plus RAG knowledge self-update (verified facts upserted into the knowledge graph) and online preference learning (rolling approval-rate signal)
• Production Data Layer (P1): A unified DataSource bus connects MES / ERP / PLM / WMS (config-driven REST connectors, env-injected) and a time-series store (in-memory ring buffer + optional InfluxDB) — agents auto-switch seed→live, closing the previous gateway/seed disconnect. Every connector degrades gracefully (no source = safe fallback to seed)
• Live Knowledge Graph & Multi-Tenant Data (P2): The knowledge graph ingests live data from the DataSource bus (periodic sync, never rewrites business numbers); each tenant can configure its own MES/ERP/WMS/time-series connections via GET/POST/DELETE /data-sources (persisted, rehydrated on restart)
• Apache-2.0 Licensed: Fully open-source, no vendor lock-in

🧩 Agent Lineup (25 Agents)

Shop-Floor Operations (11 Agents)

Enterprise Decision Brain (9 Agents)

Platform & Governance (5 Agents)

🎼 Multi-Agent Orchestration (V1.5)

The leap from "25 isolated agents" to "one collaborative team": a single goal like "improve OEE" automatically triggers 5 agents (OEE + changeover + maintenance + yield + energy) to work in parallel, then aggregates cross-domain insights into one report.

8 preset collaboration templates out of the box — just describe your goal, the platform picks the right team:

Three-tier decomposition (resilient fallback):

1. Preset templates — 8 most common scenarios, zero LLM cost
2. LLM-enhanced — for complex long-form goals, LLM picks the team
3. Keyword aggregation — always-available rule-based fallback

# Quick try (Local Python)
curl -X POST http://localhost:8000/sessions/multi-agent \
  -H "Content-Type: application/json" \
  -d '{"goal": "新产品导入评估"}'
# Returns: { session_id, plan: { sub_tasks: [...], rationale: "..." } }

curl -X POST http://localhost:8000/sessions/{session_id}/approve-multi \
  -H "Content-Type: application/json" \
  -d '{"approved": true}'
# Returns: { report: { summary, cross_findings, key_metrics, priority_actions, ... } }

See docs/MULTI_AGENT_ORCHESTRATION.md for the full design.

🚀 Quick Start

Option 1: Local Python (simplest, auto-degradation)

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env        # Fill in your LLM_API_KEY at minimum
python -m src.runtime.main
# Open http://localhost:8000/docs

No PostgreSQL or Neo4j required — the platform runs with SQLite + in-memory graph + simulated gateways; set ZHIYAN_DEMO_DATA=1 for demo data.

Option 2: Docker (full stack with PG + Neo4j + frontend)

cp .env.example .env        # Fill in your LLM_API_KEY
docker compose up -d
# Frontend: http://localhost:8080     API: http://localhost:8000

👋 如何参与

来了先别迷路 —— 四件事，任选你能做的：

• 🐛 提 Bug → Issues
• 💡 提需求 → Discussions
• 🛠 写代码 → 认领 Good First Issue
• ⭐ 用得好 → 点个 Star
• 💬 中文社区 / 教程案例 → 微信搜索公众号「工业5点0产业生态联盟」，获取上手教程、行业案例与活动

🙏 致谢 / Contributors

EvolvIQ 由社区共建，特别感谢早期朋友让这个项目有了第一批观众与更稳的代码：

• ⭐ 首批 Star：@elysium3927 @madhanio @Yangj2003 —— 你们是这个项目最早的一批观众，非常感谢。
• 🐛 @xingswxingsw —— 提交了一系列高质量 bug 报告（#44–#50：网关前缀、子路径部署、认证解包、监控告警、顶栏布局……），每一个都精准命中真实问题，平台因此更稳。欢迎继续提 issue！
• 🤝 每一位提 Issue、PR 与建议的朋友 —— 这个项目是活的，我们会持续维护。

想被写进致谢？提一个被合并的 PR，或报告一个被修复的 bug，我们就会把你加进来。

🏗 Architecture

Live Console

Key design principles:

• Deterministic by default: All agent analysis runs on seed/production data with zero LLM hallucination
• Facts are facts: Every number and action is traceable, auditable, and verifiable
• Graceful degradation: Every dependency can fail independently — platform stays up
• MCP-standardized: All 65 tools exposed via Model Context Protocol (HTTP + stdio)

📚 Documentation & Community

• 🗺 Roadmap: ROADMAP.md
• 🔒 Security: SECURITY.md
• 📝 Changelog: CHANGELOG.md
• 🏷 Release Notes: RELEASE_NOTES.md
• 📖 User Guide: docs/GUIDE.md · 中文
• 📘 Application Whitepaper: docs/WHITEPAPER.md
• 📗 Technical Whitepaper: docs/TECHNICAL_WHITEPAPER.md
• ⚡ Practicality Assessment: docs/PRACTICALITY_ASSESSMENT.md
• 🌍 Global Alignment: docs/GLOBAL_ALIGNMENT_REPORT.md
• 🏢 Enterprise Application Guide (CIO / implementation): docs/ENTERPRISE_GUIDE.zh.md
• 🛡 Risk & Governance Whitepaper (EN): docs/RISK_GOVERNANCE_WHITEPAPER.md
• 🛡 风险与治理白皮书 (ZH): docs/RISK_GOVERNANCE_WHITEPAPER.zh.md
• 📊 Datasheet & Competitive Comparison: docs/DATASHEET.md
• 🌐 Custom Domain Setup Guide: docs/DOMAIN_GUIDE.md
• 🖼 Social preview image: og_image.png (upload in repo Settings → Social preview)
• 🎬 Demo & explainer videos: see release assets / contact maintainers

🏛 治理 / Governance

EvolvIQ 是社区驱动的开源项目，治理原则公开透明：

• 方向讨论 → GitHub Discussions 与 Issues。重大变更先提 RFC（在 Discussions 开帖），再动手。
• 人在回路：任何自动进化（提示词 / 策略 / 知识图谱更新）都只进入 proposed，必须经人审批（approve → apply）才会生效——平台不会自己改业务数字。
• 决策透明：Roadmap 公开、Release Notes 公开、贡献入口公开。
• 许可证：Apache-2.0，永久全开源，无厂商锁定。

🛡 License

Apache 2.0. See LICENSE.

Built for the next generation of intelligent manufacturing. EvolvIQ is a trademark of Shanghai Dute Technology Co., Ltd.