Requirement evaluation
Read the value, the limit and the margin
Each executable requirement is evaluated against the declared design. Satisfied, violated and indeterminate results stay distinct, and missing inputs are never filled in.
Engineering workflows
Claude Code, Claude Desktop, GitHub Copilot, Codex, Cursor and ChatGPT can call MBEstudio over MCP. Your agent writes the specification; MBEstudio compiles, checks, simulates and returns results it can act on. The screenshots below are real sessions, captured from the terminal.
Agent workflows
Each step is one tool call your agent makes. Free steps never charge. Paid steps show an estimate first and stop at the spending limit you approve.
Add the server once, sign in with your account email in the browser, and ask the agent to check the connection. Here Claude Code calls describe_mbestudio, which never starts a paid job, and lists what it can now do that it could not do alone.
Real session, Claude Code 2.1, October 2026. The last paragraph of the reply, about a stale local note, is cropped.
The agent writes the system as structured data: functions, components with units, allocations, interfaces, constraints and requirements. estimate_cost runs the same checks the paid tools run and lists every problem as a blocker.
In this session the thermal requirement in the first draft pointed at an attribute that did not exist yet, on purpose, so the blocker is visible. Claude declares the attribute, validates again, and gets a 10 credit estimate before anything is charged.
It also lists the two numbers it filled in itself, a typed mass total and a placeholder case temperature, so an engineer knows what still needs evidence.
Real session, Claude Code 2.1, October 2026, uncropped.
With your approved limit, the agent runs the work it cannot do on its own: compile the specification to SysML v2, check it with the native compiler and shape checks, and execute the declared state machines. Long jobs return an ID immediately, and the result arrives as structured data.
The job appears in your studio workspace, tagged via agent, where you open the model, its source and its findings.
// estimate first, then run within the approved limit
estimate_cost({ tool: "build_full_system",
cost_mode: "agent_authored", specification })
build_full_system({ cost_mode: "agent_authored",
specification, max_spend_credits: approved_limit,
request_id })
get_job_status({ job_id })
get_job_result({ job_id })
// returns: SysML v2 source, compiler and shape findings
// with coverage, requirement verdicts over time,
// unmatched events, and the credits actually used
Why connect it
A coding agent can write SysML text and reason about a design. It cannot compile that text against a SysML v2 implementation, execute the model, or produce results that someone else can rerun. Those are the operations MBEstudio adds.
| Engineering operation | An agent alone | The agent with MBEstudio |
|---|---|---|
| Compile a system model | Writes SysML text that looks right, with no compiler to say whether it is. | generate_system_model compiles a structured specification into a SysML v2 model with requirements and interfaces. |
| Check the model | Reviews its own output with the same reasoning that produced it. | verify_model runs the native compiler and supported shape checks, and reports which areas the checks covered. |
| Execute behavior | Describes what the state machine should do. | run_simulation executes declared state machines and returns transitions, requirement verdicts over time and unhandled events. |
| Compare alternatives | Argues for a choice. | run_trade_study sweeps up to 256 declared design points against stated objectives and returns the Pareto front. |
| Produce geometry and files | Can write a file, not place parts in an assembly. | build_3d_model and export_model return envelopes or authored geometry, STL, scene JSON, engineering CSV and HTML reports. |
| Leave a record | The result lives in a chat transcript. | The model, findings and results land in your studio workspace, where an engineer can inspect and rerun them. |
Simulation is discrete. There is no fluid, rigid-body or FMU solver in the agent tools, and satisfaction depends on the equations and inputs the model declares. See the known limits.
Supported clients
MBEstudio is a remote MCP server at https://studio.vectorstreamsystems.com/mcp. Sign in with OAuth, or use an API key for scripts.
Studio workflows
The same evaluations run in the studio, where an engineer reviews a model whether a person or an agent wrote it.
Requirement evaluation
Each executable requirement is evaluated against the declared design. Satisfied, violated and indeterminate results stay distinct, and missing inputs are never filled in.
State behavior
Run the declared state machines and see when a requirement applies. Intervals where it does not apply are shown as not applicable, not counted as satisfied.
Design alternatives
Sweep declared alternatives and design variables, inspect feasible and infeasible candidates and the Pareto front, then apply one to the model.
Internal block diagrams
Typed ports and connections in the diagrams, the 3D assembly and the SysML v2 source describe the same structure.
Try it yourself
Starting points
Street Racer, performance hatchback, midsize airliner, Home Lab Rack, enduro dirt bike, cinematic quadcopter and autonomous semi truck. Open one, or point your agent at your own specification.
MBEstudio is under active development. For problems or support, email streamline@vectorstreamsystems.com.