AI for SysML v2
AI SysML v2, from requirement to checked model
Write what the system has to do in plain language. MBEstudio's agent writes the SysML v2 text, the compiler checks it and repairs what fails, and every requirement comes back with its value, limit and margin.
Worked example
Two sentences in, SysML v2 text out
What you write
"A survey drone with a 360 Wh battery and a 900 W hover draw. It shall hover for at least 20 minutes on one battery, and weigh no more than 6 kg at takeoff."
The agent turns that into parts with attributes, derived values written as expressions, and one requirement per "shall", each with a subject and a constraint the compiler can evaluate.
It is ordinary SysML v2 textual notation: a file you can read, diff, review and keep in git.
package SurveyDrone {
private import ISQ::*;
private import SI::*;
part def Battery {
attribute energy : EnergyValue = 360 [W*h];
attribute mass : MassValue = 2.1 [kg];
}
part def Airframe {
attribute mass : MassValue = 3.4 [kg];
}
part def Drone {
part battery : Battery;
part airframe : Airframe;
attribute hoverPower : PowerValue = 900 [W];
attribute takeoffMass : MassValue = battery.mass + airframe.mass;
attribute endurance : DurationValue = battery.energy / hoverPower;
}
requirement def <'SD-R1'> HoverEndurance {
doc /* The drone shall hover for at least 20 minutes on one battery. */
subject d : Drone;
require constraint { d.endurance >= 20 [min] }
}
requirement def <'SD-R2'> TakeoffMass {
doc /* The drone shall weigh no more than 6 kg at takeoff. */
subject d : Drone;
require constraint { d.takeoffMass <= 6 [kg] }
}
part drone : Drone;
satisfy HoverEndurance by drone;
satisfy TakeoffMass by drone;
}What the compiler reports
A verdict with numbers, not a green tick
This is the output of MBEstudio's compiler for the model above. Units are checked for dimensional consistency, so watt-hours divided by watts is a duration and is compared against minutes correctly.
| Requirement | Status | Value | Limit | Margin |
|---|---|---|---|---|
| SD-R1 Hover endurance | Satisfied | 24 min | at least 20 min | 4 min (20%) |
| SD-R2 Takeoff mass | Satisfied | 5.5 kg | at most 6 kg | 0.5 kg |
Delete the hover power and SD-R1 becomes indeterminate. A missing input never defaults to a pass, so a model an AI drafted from an incomplete description cannot look finished when it is not. SD-R2 still evaluates, because it does not depend on that value.
How the agent works
Generate, compile, repair, evaluate
Generate
The agent writes SysML v2 parts, ports, connections, equations and requirements from your description, or maps KiCad netlists, Onshape BOM CSV, ReqIF, MATLAB and CSV tables onto parts you already have.
Compile and repair
Every result goes through the compiler before you see it. When something fails to resolve or type-check, the agent reads the error and repairs the text, so the model you open has passed the same checks as one written by hand.
Evaluate
Requirements are evaluated against the model, across the states where they apply. Every value says whether it was stated, computed from an expression or estimated and flagged for review.
Why the text matters
SysML v2 is a systems language an AI can write like code
SysML v1 models live in a tool's database and its diagrams. SysML v2 adds a standard textual notation, so the same model is a plain file. That is what makes AI authoring practical: a language model is good at writing text, and a compiler can check text deterministically.
MBEstudio keeps that text as the single source of truth. An edit in the 3D view, a diagram or a table becomes an edit to the .sysml file. The studio, the MCP server and the command line run the same Rust compiler, built to WebAssembly, so the verdict in the browser is the verdict your agent or your CI job gets.
The compiler implements a supported subset of SysML v2 with a geometry profile, and the model reference and current limitations say exactly where that boundary is. For the wider picture, read AI MBSE for SysML v2 or the guide to AI MBSE.
Use it from your agent
SysML v2 checks inside Claude Code, Codex and Cursor
Claude Code, Codex, Cursor, Claude Desktop and ChatGPT connect over MCP to validate, evaluate, edit and simulate SysML v2 models with the same compiler the studio uses. Setup steps for each agent.
https://studio.vectorstreamsystems.com/mcp
Other tools
How other SysML v2 tools compare
Several SysML v2 tools now add AI assistants. The SysML v2 tool comparison covers thirteen of them, with where each runs, who it suits best and how MBEstudio differs.
FAQ
AI and SysML v2 questions
What is AI for SysML v2?
It is using an AI model to write, edit and check system models in the SysML v2 textual notation. Because SysML v2 models can be plain text, an AI can author them the way it writes code, and a compiler can check every result before an engineer relies on it.
Can AI generate SysML v2 from requirements?
Yes. Describe the system and its requirements in plain language and MBEstudio's agent writes SysML v2 parts, attributes and requirement definitions, compiles them and repairs what fails. Each requirement is then evaluated against the model with its value, limit and margin.
How do I know the generated model is right?
The compiler checks names, types and units before you see the model. Every value says whether it was stated, computed or estimated, and a requirement whose inputs are missing reports indeterminate instead of passing. You review the source and the requirement results together.
Does MBEstudio support all of SysML v2?
No. The compiler implements a supported subset of SysML v2 with a geometry profile. The model reference and the roadmap list what is supported and what is not, and the bundled examples show the syntax the deployed compiler accepts.
Can I use my own AI agent?
Yes. Any agent that speaks MCP, including Claude Code, Codex, Cursor, Claude Desktop and ChatGPT, can connect to https://studio.vectorstreamsystems.com/mcp and use the same compiler to validate, evaluate, edit and simulate models.
What does it cost?
The Individual plan is $39 a month and includes 2,000 AI credits a month. Credits cost $0.01 each and can also be bought as one-time packs without a subscription. Every AI job shows its maximum cost before it starts.