Appearance
Iris
Raylux Iris adds AI to the platform: an assistant for the people who build projects, short summaries of what a part of the plant is doing, a chat box for operators, agents that look into alarms, and a way for outside AI clients to read the gateway. It is for administrators who decide whether and how to turn it on, and for the engineers and operators who then use it.
It is off on every new install. Nothing is sent anywhere until an administrator adds a model, turns a feature on, and the person asking holds the right permission.
Decision support, not a safety function
What Iris writes is AI-generated and can be wrong. Do not use it in place of an operator's judgment, a plant procedure or a safety system, and do not use it to decide whether something is safe to run. Every Iris screen says so, and the model is told the same. Whether the Raylux license agreement says more about this is being settled separately; until it does, treat this paragraph as the rule.
What Iris is made of
Every Iris feature is built from the same four parts, all inside Nexus.
| Part | What it does |
|---|---|
| One provider | Nexus is the only thing that talks to a model. You choose the model: OpenAI, Anthropic (Claude), Azure OpenAI, or one you run yourself (Ollama, vLLM, LM Studio). Studio, Spectrum and Lens never call a model directly, and the key never leaves the gateway's secrets store. |
| One set of tools | The model looks things up through tools: tags, history, alarms, devices, project resources. A tool does only what the signed-in person could already do, checked on every call. |
| Proposals | The model cannot change anything. If it wants a change, it files a proposal. The change happens only when a person who holds the permission for it approves it, and it goes through the same write path as a hand edit. See Approvals. |
| Limits and a record | Token budgets and rate limits per person, per provider and per session; a transcript of each conversation; and an audit-log row for every call. See Data, limits and records. |
Iris never runs on the threads that scan devices, evaluate alarms, record history or send notifications. A slow or dead model slows Iris down and nothing else.
The surfaces
| Surface | Where you find it | Needs |
|---|---|---|
| Designer assistant | Studio: the Iris dock, and Explain, Fix, Generate and Complete in script, named-query and expression editors | AI Use; the aiCore license module |
| Summaries | ConfigHub (Summarize), Studio (Summarize with Iris), and the Iris Summary component on a Spectrum screen | AI Use; aiCore |
| Operator chat | The Iris Chat component on a Spectrum or Lens screen | AI Use; aiCore and aiOperator |
| Agents | Studio: Iris Agents. ConfigHub: Iris Agent Runs. An Iris root-cause analysis stage in an alarm notification policy | aiCore and aiAgents |
| Gateway scripts | raylux.ai.ask() in a Startup, Scheduled or Tag Change script | aiCore |
| AI clients and your own tools | The MCP endpoint, for Claude Desktop and similar | Read-only is free; your own tools and write keys need the mcp module |
Each surface has its own switch under Security › Iris AI › Features, and each is off until you turn it on. Turning Iris on walks through it.
Limits
- Iris does not write by itself. Not a chat, not an agent, not an MCP client and not a script. It files proposals; people approve them.
- A write feature needs restricted tag access. On a gateway whose tag access is still open, anyone can write any tag with no sign-in, so an approval step would protect nothing. While tag access is open Iris offers no tool that files a proposal, and it refuses to issue a key that could. Set tag access to restricted first. See Turning Iris on.
- What reaches the model leaves the gateway. With a cloud model, plant data goes to that provider. Pick a model on your own network if that is not acceptable. Data, limits and records lists what each feature can send.
- No awareness of a backup node. Nexus has no redundancy mode yet, so Iris has none either: it runs on the gateway you configured it on.
- Your own Python tools are not sandboxed against hostile code. Only people you trust should author them. See AI clients and your own tools.
Where to go next
- Turning Iris on: provider, features, limits.
- Approvals: how a proposal becomes a change, and how to undo one.
- Safety and what protects you: what Iris is built to resist, and what it is not.
- AI clients (Iris MCP): keys and connection details.