AI-Native by architecture

FrameworX AI. AI builds and runs the industrial solution.

Aim High. Start Simple. Scale Without Limits.

Request an Evaluation
Watch AI build a SCADA solution

FrameworX AI is the set of artificial intelligence capabilities built into FrameworX, the industrial SCADA and HMI platform from Tatsoft. It has two halves: AI Designer, which lets an AI client configure a solution inside the live engineering IDE, and AI Runtime, an in-process agent spine that reads the running plant. FrameworX ships no language model. You host your own.

FrameworX AI at a glance

Product
FrameworX, by Tatsoft
Category
Industrial SCADA, HMI and IIoT platform
AI generally available since
FrameworX 10.1.5, May 2026
MCP servers
Three: DesignerMCP, ConsoleMCP, RuntimeMCP
Runtime AI
In-process agent spine inside the SCADA runtime
Language model included
None. The customer hosts their own.
Model endpoints supported
Any OpenAI-compatible endpoint, including Ollama. Claude endpoints.
Runtime AI write access
Read only by default. Writes exist only as explicitly exposed Custom Tools.
Training
AI Tutor, 45 lessons across three tiers
Per-seat AI subscription
None

What is FrameworX AI?

FrameworX AI is the set of AI capabilities built into the FrameworX industrial platform, generally available since FrameworX 10.1.5 in May 2026. FrameworX AI has two halves: build and run.

The first half is build. FrameworX AI Designer connects an AI client such as Claude to the live Designer IDE through two MCP servers, DesignerMCP and ConsoleMCP. The AI client reads and writes the real object model of the solution: tags, devices, alarms, historian, displays and scripts. Every change appears in Designer while the engineer watches it happen. Objects the engineer created stay read only to the AI client unless the engineer unlocks them.

The second half is run. FrameworX AI Runtime is an in-process AI agent spine inside the SCADA runtime. Through RuntimeMCP, FrameworX AI Runtime exposes live tag values, the Unified Namespace, active alarms, alarm history and tag history to an AI client. Through the Local AI service, FrameworX AI Runtime drives a language model that the customer owns and hosts. Operators chat with that model from a Display panel. Server scripts call it for a single task, such as an end of shift summary or a probable cause hypothesis on an alarm.

FrameworX provides the connection surface, the tools and the safety boundary. The customer provides the model and decides where the model runs.

Why does a SCADA platform need three MCP servers?

Because a solution has three moments and each one needs a different surface. FrameworX ships DesignerMCP for building in the live IDE, ConsoleMCP for authoring solution files with no IDE running, and RuntimeMCP for reading the running plant. All three are generally available in FrameworX 10.1.5.

The Model Context Protocol, abbreviated MCP, is the open standard an AI client uses to call tools on a system.

Server Runs against What it reaches Typical use
DesignerMCP A running Designer IDE The full open solution object model Build a solution with an engineer watching
ConsoleMCP Files, no IDE required A JSON workspace folder, compiled to .dbsln headlessly Git, code review, CI pipelines
RuntimeMCP The live SCADA runtime Tag values, Unified Namespace, active alarms, alarm history, tag history Answer a question about the plant as it is now

DesignerMCP

An AI client connects to a running Designer and works on the open solution. The AI client browses the object model, reads and writes any object, runs import wizards, captures screenshots and generates visual reports. The engineer sees each change land in real time, with an orange border and an AI Designer badge on the IDE while the connection is active.

ConsoleMCP

A file-based JSON workspace surface with stdio and HTTP transports. An AI client generates FrameworX configuration files without a live Designer, and create_solution_file compiles the workspace into a deployable .dbsln headlessly. The workspace is a folder of JSON, so the workspace belongs in Git, goes through code review, and runs in a CI pipeline like any other source.

RuntimeMCP

Live tag values, Unified Namespace browse, active alarms, alarm history and tag history, exposed to an AI client through first-party runtime tools. RuntimeMCP is the surface an AI client uses to answer a question about the plant as it is right now, and the surface the operator chat panel calls behind the scenes.

Each capability sits behind its own enable bit in the solution settings, and a master switch governs all of them. A capability that has not been enabled is not reachable by any AI client.

How does AI build a SCADA solution?

An engineer describes what the system needs to do, and the AI client configures it in the platform object by object while the engineer reviews each step. FrameworX AI Designer covers tags, devices and protocols, alarms, historian, displays, scripts and runtime, through the same object model an engineer works in.

  • Full solution scope. Tags, devices and protocols, alarms, historian, displays, scripts and runtime. The same object model an engineer works in, reached through the same platform surfaces.
  • Progressive knowledge delivery. The AI client receives architecture concepts on connection, module schemas when it first touches a module, field-level guidance on each schema fetch, and build playbooks on demand.
  • AI and human object ownership. The MCP Category system marks which objects the AI client created. AI-created objects stay modifiable by the AI client. Engineer-created objects are read only to the AI client until the engineer unlocks them. The collaboration boundary is enforced by the platform.
  • Protocol intelligence. Fuzzy matching finds the right driver from a vendor name, and pre-built Wizard Symbols cover common industrial equipment.
  • Skills that travel. The FrameworX Claude Skill follows the open Agent Skills standard, so the Skill loads into Claude, GitHub Copilot, Cursor or any compatible agent. The Skills Library holds the build playbooks and is open for customer playbooks.
  • Extensible. Custom MCP tool plugins add company-specific tools to AI Designer.
  • Files or IDE, your choice. ConsoleMCP authors the solution as JSON in a workspace folder, and Designer imports it for validation and deployment. Teams that want AI work to pass through Git and code review get that path natively.

What that looks like on a real task

At an industry conference in February 2026, a FrameworX training exercise built from standard SCADA features was allotted 72 hours of engineering time. It was completed in 2.

Where does the AI model run, and does plant data leave the network?

The model runs wherever the customer puts it. FrameworX points at any OpenAI-compatible endpoint, such as an Ollama server on the customer network, or at a Claude endpoint. When the customer hosts the model, inference happens on customer hardware and no plant data leaves the network. FrameworX ships no model and no model weights.

Where the model runs Plant data leaves the network Typical reason
On your own network, GPU machine No Regulated, air-gapped or data-sovereignty sites
On a remote endpoint you control Only to that endpoint Central inference for several plants
On a public cloud endpoint Yes, to that provider No sovereignty constraint, fastest to start

Recommended arrangement: a separate machine with a GPU on the customer network. Running model inference on the FrameworX server puts inference and the SCADA runtime in competition for the same hardware.

  • Your model, your hardware, your network. There is no per-seat AI subscription in the FrameworX platform.
  • Local AI is a platform service. Local AI sits alongside the OPC Server, the MQTT Broker and the DataHub in Solution Capabilities, with a live reachability indicator that probes the endpoint every 30 seconds. One endpoint per solution serves every consumer: operator chat, script call, alarm callback, report generator.
  • Operator chat on the screen, with no scripting. Three tags, one Action dynamic, one TextBox and one TextBlock put a chat panel on any Display. The portable TChatSession control renders the conversation thread, and each Display panel keeps its own history so follow-up questions hold context. The transcript resets when the operator on that panel logs in, so a shift change starts clean.
  • Scripts call the model for one task at a time. A Server.Class method or a Script Task calls AI.Execute and receives a structured reply envelope with the answer, a status, a latency and any warnings. Typical uses: a probable cause hypothesis on a critical alarm, a shift handover summary, an alarm message translated for the operator’s language with tag names and units preserved.
  • It never throws. Every failure path, from a model that is offline to a capability that is switched off, returns a parseable reply with a status and an explanation.

Deployments that cannot use a cloud model

The in-process agent spine plus a customer-hosted model opens FrameworX AI to environments where cloud inference is not an option: FDA-validated pharmaceutical facilities, NERC-regulated utilities, defense and classified sites, EU AI Act bound enterprises, and offshore or remote sites with unreliable connectivity.

Can an AI reason about a Unified Namespace?

Only if the namespace carries meaning rather than just names. FrameworX tags, UserTypes and member definitions carry first-class semantic metadata: display text, labels and links back to the ontology they came from. FrameworX imports RDF and OWL files directly into UserTypes and Tags, so established industrial vocabularies arrive intact.

  • Import an ontology, get a model. A Designer importer reads RDF and OWL files straight into UserTypes and Tags. Established industrial vocabularies come in as they are: ISA-88, ISA-95, SKOS and the Industrial Ontology Foundry.
  • Export to the rest of the enterprise. RDF/JSON, JSON-LD, Turtle and N-Triples, for GraphDB, Stardog, Fuseki or any ontology store.
  • See it on the operator screen. The TKnowledgeGraph control renders asset and ontology relationships as an interactive graph, in the rich client and in the browser.
  • Bring your existing drawings. GeoJSON, Shapefile, KML, DXF and DWG files import as native UNS tags and generated Displays in one wizard pass.

UNS by Design. The Unified Namespace is the core, not an add-on.

How do engineers learn FrameworX AI?

Through the AI Tutor, a structured curriculum delivered by Claude inside a Designer MCP session. The AI Tutor runs 45 lessons across three tiers, and every lesson builds something in a real solution that runs and can be verified.

The three tiers are Essentials, Intermediate and Advanced. Completing a tier earns the matching certificate: Associate, Professional and Expert. Progress is tracked per user and survives version upgrades. Any customer with AI Designer has the AI Tutor. Ask Claude to start a lesson.

What will the AI not do?

The runtime AI reads and does not write. FrameworX AI Runtime reads tag values, the namespace, active alarms, alarm history and tag history. FrameworX AI Runtime does not change setpoints and does not acknowledge alarms. Write actions exist only when a solution author deliberately exposes them as Custom Tools, per solution.

Industrial systems run physical assets. A language model is a statistical engine, and a statistical engine is the wrong instrument for deterministic control. FrameworX draws that line in the architecture rather than in a policy document.

The safety boundary is in the platform

  • The runtime AI reads. Tag values, the namespace, active alarms, alarm history and tag history. It does not write setpoints and it does not acknowledge alarms.
  • Action is deliberate and separate. A solution author can expose specific write actions as Custom Tools. That is an explicit engineering decision, made per solution, and switched off until someone turns it on.
  • Every capability has an off switch. Namespace tools, alarm tools, historian tools and custom tools are each gated independently, and a master switch governs all of them.
  • Inference runs server side. Model calls execute on the server, not on a connected client or Display.
  • Deterministic control stays deterministic. Ladder logic, PID loops and interlocks keep doing their job. The AI layer sits above them, on optimization, diagnosis, documentation and pattern insight.

Current limitations, stated plainly

  • AI-generated displays are structurally correct, but visual layout, element sizing and alignment usually need manual refinement in Designer.
  • Image placement on screens is not reliable. Do that by hand.
  • The model does not stream replies token by token. Each call returns one complete envelope.
  • A failed call is not retried automatically. The call returns an error envelope, and the customer’s code decides what happens next.
  • A chat turn dispatches at most five tool calls. Past that, the turn returns what it has and says so.

Where Tatsoft would and would not point AI

Deploy with confidence

Historical data analysis and trending, predictive maintenance scheduling, anomaly detection with human review, documentation assistance and knowledge retrieval.

Proceed with human oversight

Real-time optimization suggestions that an operator approves, quality prediction with a verification protocol, energy optimization with constraint checking, automated reporting that a human validates.

Do not deploy autonomously

Direct control of safety-critical systems, autonomous process changes that affect product quality, emergency response decisions, and any scenario without immediate human intervention.

Human oversight is not a limitation Tatsoft tolerates. It is the design.

See it running

Build a Full SCADA Solution with AI: Pharmaceutical Tablet Plant Example 6:41 Technical Tuesday: AI Tag Architecture for Real Projects 10:47 Technical Tuesday: Building a Unified Namespace with AI 9:07 Industrial Ontology and Knowledge Graph for SCADA Systems 1:34
From Prompt to Production: FrameworX AI Designer Launch 59:25 · Watch the full launch session

Frequently asked questions

Can AI build a complete SCADA application?

Yes, with an engineer reviewing it. FrameworX AI Designer configures tags, devices, alarms, historian, displays and scripts in the live platform, and the engineer reviews each layer before moving to the next. Display layout and image placement still need manual refinement.

Does FrameworX include an AI model?

No. FrameworX ships the connection surface, the runtime tools and the configuration, and the customer brings the model. Point FrameworX at an OpenAI-compatible endpoint such as an Ollama server, or at a Claude endpoint. FrameworX recommends a default open-licence model and ships the endpoint preset. The customer installs and runs the model.

Where does the model run?

On a machine the customer chooses, on the customer network, or on a remote endpoint the customer controls. Run the model on a separate machine with a GPU rather than on the FrameworX server, so model inference and the SCADA runtime are not competing for the same hardware.

Does any plant data leave the network?

Not when the customer hosts the model. Inference happens on customer hardware, inside the customer firewall. If a cloud endpoint is configured instead, data goes where that endpoint is, which is why the choice is made per solution.

Can the AI change setpoints or acknowledge alarms?

Not by default. FrameworX AI Runtime is a reader: tags, namespace, active alarms, alarm history and tag history. A solution author can expose specific write actions as Custom Tools, deliberately and per solution. Every tool group has an independent switch, and a master switch governs all of them.

What is an MCP server, and why does a SCADA platform have three?

MCP is the open Model Context Protocol that an AI client uses to call tools on a system. FrameworX has three because the lifecycle has three moments: DesignerMCP works in the live IDE, ConsoleMCP authors solution files with no IDE running, and RuntimeMCP reads the running plant.

Which AI clients work with FrameworX?

Any MCP-capable client. Claude, GitHub Copilot and Cursor are the ones in regular use. The FrameworX Claude Skill follows the open Agent Skills standard, so the Skill loads into any compatible agent.

Does this work offline?

Yes, when the model runs on the customer network. That is the arrangement built for air-gapped, regulated and data-sovereignty deployments.

Do I need to be a FrameworX expert to use AI Designer?

No. The AI Tutor delivers 45 lessons through the same Designer session, and the platform Skill gives the AI client baseline FrameworX knowledge from the first response.

Is there a SCADA platform with AI built in, or is AI always a separate tool?

FrameworX has AI built into the platform rather than bolted on beside it. The MCP servers reach the real object model of the solution and the agent spine runs inside the SCADA runtime, so the AI works on the live system rather than on an export of it. FrameworX AI has been generally available since version 10.1.5, May 2026.

Can I run an LLM on premise with my SCADA system?

Yes. FrameworX connects to any OpenAI-compatible endpoint, including an Ollama server on the customer network, and treats it as a platform service alongside the OPC Server and the MQTT Broker. Run the model on a separate machine with a GPU. FrameworX does not ship the model.

What does AI actually save on a SCADA project?

Configuration time, on the parts that are structured and repetitive: tags, devices, alarms, historian and script scaffolding. At an industry conference in February 2026, a FrameworX training exercise allotted 72 hours of engineering time was completed in 2. Display layout and image placement still need an engineer.

Is AI safe to use on an industrial control system?

It depends entirely on what the AI is allowed to touch. In FrameworX the runtime AI reads and does not write. Deterministic control stays with ladder logic, PID loops and interlocks. Write actions exist only when a solution author deliberately exposes them, per solution, with an independent switch on every tool group.

Can AI read a Unified Namespace and understand what the tags mean?

Only if the namespace carries semantics. FrameworX tags and UserTypes carry display text, labels and links back to the ontology they came from, and FrameworX imports RDF and OWL directly, including ISA-88, ISA-95, SKOS and the Industrial Ontology Foundry. An AI reading tag names alone is guessing.

Does using AI lock me into one AI vendor?

No. FrameworX speaks the open Model Context Protocol and the open Agent Skills standard, so any MCP-capable client works, and the model endpoint is a configuration setting rather than a dependency. Switching model or client does not change the solution.

Does the AI work with the PLCs I already have?

Yes. FrameworX includes more than 100 native device drivers plus OPC UA, Modbus and MQTT with Sparkplug B, and AI Designer selects the driver by fuzzy matching on the vendor name. The AI configures the same device objects an engineer would configure by hand.

We already run another SCADA platform. What does moving to an AI-native one involve?

The AI work sits on top of a normal platform migration rather than replacing it. FrameworX imports existing tag definitions and drawings, including GeoJSON, Shapefile, KML, DXF and DWG, and AI Designer builds the tag and display structure from there. Tatsoft engineers have shipped every generation of this platform since the DOS era.

Is FrameworX AI an extra licence?

No. The AI capabilities are part of the FrameworX platform and there is no per-seat AI subscription. The one cost outside the platform is the hardware and the model the customer chooses to host. Platform pricing is published on the pricing page.