Governed, vendor-neutral AI ops for OT / 工控 — audit, budget, undo, and graduated approval built into every tool. Read-first, because the plant floor is exactly where you want an agent on a tight leash.

English · 中文

AI agents are great at industrial operations — right up until one issues a write to a live PLC. Industrial-AIOps is a vendor-neutral OT data tap + cross-protocol troubleshooting layer for AI agents, wrapped in a governance harness so an agent's actions are:

• Audited — every operation logged to a local SQLite trail (who / what / why / when), secret-redacted.
• Bounded — per-process token/call budget + a runaway-loop breaker.
• Reversible — write/command tools record an inverse undo token (the before-value/state is captured).
• Graduated — risk tiers gate writes; the few OT-dangerous commands are off by default (dry_run), need a double-confirm, and a recorded approver (MOC).
• Read-first — the vast majority of tools only read; writes are the rare, heavily-gated exception.

Unlike the broader IT-ops family (AIops-tools), this line is a monorepo with a shared core + per-protocol connectors, and a menu-configurable MCP — a site exposes only the 1–2 protocols it runs.

Everything software-verifiable is verified (in-process servers, real protocol libraries, container loopbacks); what remains on the honest 待核实 list only real equipment can answer — physical Modbus-RTU (RS-485), EtherCAT slaves, HART gateways, live BACnet HVAC, domestic PLCs (汇川/信捷), live PLCnext, live Omron FINS PLCs, IO-Link masters, substation RTUs/IEDs. Verified-equipment reports get credited in the support matrix; features are co-designed via Issues/Discussions. Start here → industrial-aiops#28 — Call for field-testing partners (v0.10.0) (pinned).

One package, one menu-configurable MCP server. Install only the protocols a site runs (iaiops[opcua], or an edition bundle iaiops[fab]); expose only those at runtime (IAIOPS_MCP=fab). Config + the audit/undo/policy store live under ~/.iaiops/; credentials are kept in an encrypted store (secrets.enc, Fernet + master password) — never plaintext.

pip install "iaiops[fab]"     # fab = secsgem + opcua + s7 + modbus
iaiops init                   # interactive setup (encrypts credentials)
iaiops doctor                 # verify connectivity (classified, not raw errors)
IAIOPS_MCP=fab iaiops mcp     # run as an MCP server, fab profile

Since 0.10.0 there is no default tool exposure: a bare iaiops-mcp prints the selection menu and exits — pick a profile (IAIOPS_MCP=fab|factory|process|building|water|warehouse|clinical|renewables|plcnext|brain|…) or a pre-scoped iaiops-mcp-<name> entrypoint. Multi-process sites run 1 iaiops-mcp-brain + N protocol servers with IAIOPS_MCP_NO_BRAIN=1.

Also on the MCP Registry (io.github.industrial-aiops/iaiops, io.github.industrial-aiops/iaiops-energy), PyPI, and ClawHub.

Edge-native by design. iaiops rides on a hardened, centrally-managed edge host as a portable, governed edge application — mapping onto the Margo edge-interoperability roles (immutable host · compliant orchestrator · iaiops = the OT-domain app), with an optional on-box LLM brain for a fully air-gapped diagnostic path. Roadmap ⏳: container image + Margo application description + a published conformance result — not Margo-compliant yet (see the repo's docs/MARGO-ALIGNMENT.md).

⚠️ Honest validation status — pure analysis, OPC-UA, Modbus-RTU (software serial), BACnet/IP reads, TDengine/IoTDB, DNP3 and IEC-61850 monitor paths are verified against real servers/libraries/containers; live field hardware largely stays 待核实 (see each repo's README). Read-first; never write to a production control system without authorization.

AI agent 做工业运维很在行——直到它对在产 PLC 下了一次写。Industrial-AIOps 是给 AI agent 用的厂商中立 OT 数据 tap + 跨协议智能排查,外包一层治理 harness,让 agent 的动作:

• 可审计 —— 每次操作落本地 SQLite 流水（谁/做了什么/为什么/何时），密钥脱敏。
• 有预算 —— 每进程 token/调用预算 + runaway 熔断。
• 可回滚 —— 写/命令工具记录反向 undo token（捕获改前值/状态）。
• 分级审批 —— risk-tier 把写操作分级；少数 OT 危险命令默认关（dry_run）、需双重确认 + 记录审批人（MOC）。
• 只读优先 —— 绝大多数工具只读;写是罕见且重度受控的例外。

与偏 IT 的 AIops-tools 不同,这条线是共享 core + 按协议 connector 的 monorepo,且 MCP 可菜单配置——现场只暴露它实际跑的 1–2 种协议。

软件里能验证的都验证了(in-process 服务器、真实协议库、容器 loopback);诚实的 待核实 清单只有真设备能回答——物理 Modbus-RTU(RS-485)、EtherCAT 从站、HART 网关、在线 BACnet 楼宇设备、国产 PLC(汇川/信捷)、真机 PLCnext、欧姆龙 FINS 真机、IO-Link 主站、变电站 RTU/IED。经你验证的设备会署名写进支持矩阵;功能通过 Issues/Discussions 共创。参与入口 → industrial-aiops#28 — 招募现场测试伙伴 (v0.10.0)(置顶)。

一个包、一个可菜单配置的 MCP server。现场只装它跑的协议（iaiops[opcua],或行业捆绑 iaiops[fab]）;运行时只暴露这些(IAIOPS_MCP=fab)。配置与审计/undo/策略存于 ~/.iaiops/;凭据放加密存储(secrets.enc,Fernet + 主密码),绝不明文。

pip install "iaiops[fab]"     # fab = secsgem + opcua + s7 + modbus
iaiops init                   # 交互式设置(加密凭据)
iaiops doctor                 # 连通自检(给归因结论,不是裸错误)
IAIOPS_MCP=fab iaiops mcp     # 以 fab profile 跑 MCP server

0.10.0 起无默认工具暴露:裸 iaiops-mcp 打印选择菜单后退出——用 IAIOPS_MCP=fab|factory|process|building|water|warehouse|clinical|renewables|plcnext|brain|… 选 profile,或直接用预置入口 iaiops-mcp-<name>。多进程站点:1 个 iaiops-mcp-brain + N 个带 IAIOPS_MCP_NO_BRAIN=1 的协议 server。

同时在 MCP Registry(io.github.industrial-aiops/iaiops、io.github.industrial-aiops/iaiops-energy)、PyPI、ClawHub 上架。

边缘原生定位。 iaiops 以边缘应用(edge application)的形态跑在加固、集中管理的边缘主机上——对齐 Margo 工业边缘互操作标准的角色划分(不可变主机 · 合规编排器 · iaiops = OT 域应用),并可对接本机 LLM 脑实现完全离线(气隙)的诊断路径,数据不出厂。Roadmap ⏳:容器镜像 + Margo application description + 公开的 conformance 结果——目前尚未 Margo-compliant(见仓库 docs/MARGO-ALIGNMENT.md)。

⚠️ 诚实验证状态 —— 纯分析、OPC-UA、Modbus-RTU(软件串口)、BACnet/IP 读、TDengine/IoTDB、DNP3 与 IEC-61850 监视路径已对真实服务器/库/容器验证;真实现场硬件大多仍 待核实(见各仓库 README)。只读优先;未经授权勿对生产控制系统写入。

MIT licensed. Community-maintained. Not affiliated with, endorsed by, or sponsored by the vendors or standards bodies of the systems these tools operate (Siemens, Rockwell Automation, Mitsubishi Electric, the OPC Foundation, SEMI, the Eclipse Foundation, etc.); all trademarks belong to their respective owners.