STREAMSIGHT
Early Preview · pre-1.0
Your diagrams are the model. No separate source of truth.
Industrial teams already map their plants visually: engineers build hundreds of dashboards in tools like PI Vision that lay out equipment, relationships, and live process data. Streamsight turns that visual context into a machine-readable process graph, a digital representation of how the plant actually operates. Then it puts an engineering agent on that graph. Describe a symptom and it traces the process, runs real analysis, and builds a live dashboard to show what it found, bringing the AI productivity that reshaped software development to industrial operations.
The engineering agent
Describe a symptom in plain language. The agent traces the dependency graph to find what feeds the trouble, pulls the relevant tag history, and works toward a root cause, naming every tag and entity by its real id so you can check its work.
The agent runs a genuine Python sandbox: correlations, regressions, change-point detection, spectral analysis. It installs whatever analytical package the question needs, so the answer is computed from your data, not guessed.
When the analysis lands, the agent assembles a live dashboard: KPI tiles and trend charts wired straight to your tags, self-updating on their own timers. The finding isn't a paragraph you forget, it's a view you keep.
You stay in control
The agent can extend the model itself: declare a new tag, add a piece of equipment, drop a node onto the diagram, write a status threshold. Nothing lands silently.
Why it's different
The hard part of industrial AI has always been context, and the usual answer is to stand up a separate data platform and run a long project to mirror your plant inside it. Streamsight takes its context from where it already lives: the operational views engineers build and maintain every day. Keep the views current and the agent's model stays current with them.
Build flows visually. Nodes are the steps, equipment, or services; edges show how they connect. A shared process library keeps reusable shapes, tags, and edge types consistent across diagrams.
Attach a data item to any node: a tag from your historian, a synthetic source for development, or a derived expression that combines other bound values into a new quantity. The diagram updates live. PI Web API is the first integration that's shipped, with more on the way.
Status profiles turn thresholds into color. Operators see green / amber / red across the whole flow without reading every number, and the agent reads the same state when it decides where to look first.
Any node can open into its own sub-diagram, or reference another diagram in the library. The top-level view stays readable; the detail is one click away, when one symbol represents a whole skid, unit, or subsystem.
Open platform
Everything the in-product agent can do is a typed tool in a shared catalog, and that catalog is published over MCP. Point your own model at it, or attach an external agent like Claude Code or Cursor. The same governed tools, the same propose-then-approve safety, whoever's driving.
FAQ
An AI engineering agent for industrial operations. Engineers already map their plants visually to monitor them; Streamsight turns those diagrams into a machine-readable process graph, then puts an agent on it to diagnose problems, run real analysis, and build dashboards. Every change the agent makes is a proposal you approve.
Describe a symptom and it traces the process graph, pulls the relevant tag history, and runs real analysis in a Python sandbox. It builds a live dashboard to present the finding, and can extend the model itself: declaring tags, adding equipment, dropping nodes on a diagram, writing status rules.
No. Streamsight derives its context from the process diagrams engineers already build to monitor operations, turning those views into a machine-readable graph. There's no separate canonical data layer to stand up or keep in sync; maintaining the view maintains the model.
Every write routes through propose-then-approve: the agent submits a proposal and you approve it in the UI before anything lands. For multi-step changes it outlines a plan and waits for your go-ahead. Corrections you give it persist across conversations.
Yes. The agent runtime is pluggable, and the same typed tool catalog the in-product agent uses is published over MCP. Point your own model at it, or attach an external agent such as Claude Code or Cursor, with the same propose-then-approve safety.
Mainstream industrial historians and process data protocols. PI Web API is the first integration that's shipped; additional historians and protocols are in active development. Nodes can also bind to derived expressions or a synthetic source for development.
On-prem and single-tenant. Installed in your environment, behind your network perimeter. Single sign-on via Microsoft Entra ID (Azure AD) is supported. Streamsight is pre-1.0 and currently shown as an early preview to interested teams.
Design partner program
Streamsight is looking for teams on the cutting edge of industrial operations. Teams that want an agent working alongside their engineers, and want to be ready for where the field is headed. As a design partner, you get free access throughout the pilot, a direct line to the team building the product, meaningful influence over what gets prioritized next, and a version of Streamsight shaped around how your team actually works.
