AI Agents & MCP Workflows

Large language models excel at generating structure and exploring design spaces, but when asked to write differential equations, power-grid topologies, or safety-critical requirements in free text, they frequently hallucinate mismatched units, structurally singular DAEs, or violated physical conservation laws.

ModelScript closes this loop through @modelscript/mcp — a deterministic Model Context Protocol (MCP) server exposing 52 polyglot compilation, simulation, CAD, TGG, and formal verification tools (plus 70+ domain-specific tool extensions). Instead of guessing physics, an AI agent iteratively compiles, simulates, verifies, and calibrates models against mathematical solvers until every requirement proof passes.

---

1. Connecting Your AI Agent

@modelscript/mcp supports three transport topologies: local stdio, remote HTTP Server-Sent Events (SSE), and in-browser WebAssembly (PolyglotMcpHost).

VS Code (stdio & @modelscript Copilot Participant)

In VS Code (v1.99+), register @modelscript/mcp in .vscode/mcp.json using the "servers" root key:

{
  "servers": {
    "modelscript": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@modelscript/mcp"]
    }
  }
}

Or add it directly from your terminal via the VS Code CLI:

code --add-mcp '{"name":"modelscript","command":"npx","args":["-y","@modelscript/mcp"]}'

Installing the official ModelScript VS Code Extension (ext install modelscript.modelscript) also registers all 52 tools natively into vscode.lm.registerTool, unlocking the @modelscript GitHub Copilot Chat participant (/explain, /fix, /simulate, /verify, /optimize, /diagram) and the dedicated ModelScript AI sidebar panel.

Claude Code (stdio)

Register the MCP server directly in your project or user scope:

claude mcp add modelscript -- npx -y @modelscript/mcp

Cursor, Windsurf & Claude Desktop (stdio)

Add @modelscript/mcp to .cursor/mcp.json, ~/.codeium/windsurf/mcp_config.json, or claude_desktop_config.json using the "mcpServers" root key:

{
  "mcpServers": {
    "modelscript": {
      "command": "npx",
      "args": ["-y", "@modelscript/mcp"],
      "env": {
        "MODELSCRIPT_API_URL": "http://localhost:3000"
      }
    }
  }
}

Hosted Enterprise Hub (SSE)

For cloud agents, CI review bots, and multi-tenant teams, connect to the ModelScript Hub's authenticated SSE gateway (/api/v1/mcp/sse) using a user JWT or scoped ms_bot_* service token:

{
  "mcpServers": {
    "modelscript-cloud": {
      "url": "https://hub.example.com/api/v1/mcp/sse",
      "headers": {
        "Authorization": "Bearer ms_bot_live_9f8e7d6c5b4a..."
      }
    }
  }
}

Grounding Context Windows with /llms.txt

When prompting an LLM outside of MCP (or seeding system prompts for custom agents), point the model at https://modelscript.org/llms.txt. It is generated directly from the compiler's CLI definitions, language registry, and linter catalogs during every documentation build.

---

2. Core Agentic Engineering Workflows

Rendering diagram...

Workflow A: Autonomous Closed-Loop Physical Synthesis

Rather than writing a Modelica controller in one shot and hoping it stabilizes the plant, instruct the agent to call modelscript_closed_loop_synthesize. This single tool orchestrates a five-stage compiler/solver loop:

1. GLR Parse & Semantic Lint (modelica_lint across 307 Modelica rules)
2. Arena DAE Flattening & Pantelides Index Reduction (modelica_flatten)
3. Time-Domain Integration (modelica_simulate via SUNDIALS CVODE/IDA or Radau IIA)
4. Signal Temporal Logic (STL) Verification (sysml2_falsify_requirements)
5. Counterexample & Sensitivity Tuning: Evaluates STL robustness $\rho$ and tunes designated control gains (tunableParameters) via IMC-PID and sensitivity refinement until all five verification gates pass.

Example MCP Tool Call:

{
  "name": "modelscript_closed_loop_synthesize",
  "arguments": {
    "modelName": "ActiveSuspension",
    "modelicaSource": "model ActiveSuspension\n  parameter Real k_s = 18000;\n  parameter Real c_d = 1450;\n  ...\nend ActiveSuspension;",
    "sysmlSource": "requirement def <'REQ-RIDE-01'> PeakAccelerationBound {\n  require constraint { maxVerticalAccel <= 1.8 }\n}",
    "tunableParameters": ["Kp", "Ki", "Kd"],
    "stopTime": 5.0,
    "maxRepairIterations": 5
  }
}

---

Workflow B: Differentiable Parameter Calibration Against Test Data

When you have experimental bench data or field telemetry in CSV format, the agent can call modelica_diff_calibrate to fit unknown physical parameters (such as friction coefficients, thermal resistances, or battery internal impedances) using forward-mode Automatic Differentiation (DualNumber) paired with L-BFGS or Adam:

{
  "name": "modelica_diff_calibrate",
  "arguments": {
    "modelName": "LithiumIonCell",
    "sourceCode": "model LithiumIonCell ... end LithiumIonCell;",
    "targetCsv": "time,v_terminal\n0.0,4.18\n10.0,4.09\n20.0,4.02\n...",
    "parameters": [
      { "name": "R_internal", "initial": 0.05, "min": 0.005, "max": 0.25 },
      { "name": "C_dl", "initial": 1200.0, "min": 100.0, "max": 5000.0 }
    ],
    "optimizer": "lbfgs",
    "maxIterations": 60,
    "tolerance": 1e-5
  }
}

After calibration converges, ask the agent to invoke modelica_sensitivity to inspect local identifiability and confidence intervals around the calibrated operating point.

---

Workflow C: Bidirectional SysML v2 $\leftrightarrow$ Modelica TGG Synchronization

When a systems engineer adds a port or updates a mass budget in .sysml while a controls engineer edits equations in .mo, the digital thread can diverge. AI agents resolve cross-domain drift deterministically using three TGG & Theory Coordinator tools:

1. validate_system_consistency: Cross-checks .sysml and .mo models (or runs the 13 Nelson-Oppen Domain Theory Oracles) to detect port topology mismatches, unit conflicts, and divergent shared parameter slots.
2. tgg_diagnose_conflict: Runs static Critical Pair Analysis (CPA) and inspects concurrent edits on Triple Graph Grammar correspondence nodes.
3. tgg_reconcile_slot: Applies a formal resolution policy (source-wins, target-wins, physics-simplex, or interval_intersection) and rewrites both ASTs losslessly.

{
  "name": "tgg_reconcile_slot",
  "arguments": {
    "correspondenceId": "corr_ThermalCoolantLoop_maxPressure",
    "strategy": "interval_intersection",
    "sysmlPath": "architecture/ThermalSystem.sysml",
    "modelicaPath": "models/ThermalCoolantLoop.mo"
  }
}

---

Workflow D: Requirements Traceability Matrix (RTM) & Suspect Link Auditing

Before cutting a release or generating an FDA / ISO 26262 Design History File (msx dhf export), an AI agent can audit the workspace's Requirements Traceability Matrix:

1. rtm_query_matrix: Queries all requirements, satisfying architecture/physics elements, verification test cases, coverage percentages, and suspect links (where a downstream .mo equation or .step CAD solid changed after the requirement was last verified).
2. rtm_diagnose_suspect: Computes the structural CST/DAE delta across the suspect link and recommends whether to re-run verification (sysml2_verify) or update the requirement bound.
3. rtm_export_reqif: Exports the synchronized traceability matrix to OMG ReqIF 1.2 (.reqif) for enterprise ALM import.

---

Workflow E: Cross-Domain CAD/FEA Sizing, Legacy B2B Certification & Formal Safety

Beyond 1D system dynamics, @modelscript/mcp provides five cross-domain certification tools:

• cad_fea_closed_loop_size: Couples parametric 3D OpenSCAD/STEP geometry with tetrahedral FEA von Mises stress analysis and 3D BVH collision clearance checks, iteratively sizing ribs, walls, and fillets until structural safety factors hold.
• polyglot_migrate_and_verify_b2b: Transpiles legacy Simulink (.slx), SPICE (.cir), or Adams (.adm) models into balanced Modelica and SysML v2, then executes an ISO 26262 Part 6 Back-to-Back (B2B) trajectory equivalence proof with C99 SHA-256 codegen parity.
• ssp_diagnose_cosim_stability: Detects direct-feedthrough algebraic loops across multi-FMU SSP topologies, computes the Zero-Order Hold (ZOH) spectral radius $\rho (J_{\mathrm{ZOH}})$, and synthesizes FMI directional-derivative Newton relaxation schedules.
• formal_verify_safety_and_mcdc: Proves unbounded safety invariants via IC3/PDR and Sum-of-Squares (SOS) barrier certificates, and synthesizes DO-178C / ISO 26262 MC/DC independence-pair test vectors.
• twin_discover_sindy_and_rom: Runs Sparse Identification of Nonlinear Dynamics (SINDy STLSQ) over hardware telemetry residuals to discover unmodeled physics terms and calibrate a real-time Reduced-Order Model (ROM).

---

3. Recommended Prompt Patterns for Engineers

Use these prompt templates with Claude Code or Cursor when working in a ModelScript workspace:

** Synthesis & Verification Prompt:** "Create a Modelica model DC_BuckConverter.mo and a companion BuckConverterReq.sysml specifying a $12\text{ V}\to 5\text{ V}$ step-down converter with output voltage ripple $\le 50\text{ mV}$ and settling time $\le 4\text{ ms}$. Run modelica_lint, simulate with modelica_simulate, and iterate using modelscript_closed_loop_synthesize until sysml2_verify proves both requirements SAFE."

** Legacy Migration Prompt:** "Import legacy/PitchActuator.slx into models/PitchActuator.mo using polyglot_migrate_and_verify_b2b, run modelica_lint to resolve any diagnostic warnings, and verify that the step response matches the Simulink reference trajectory within ${10}^{-4}$ relative error."

** Grid Compliance Prompt:** _"Run pscad_lint and pscad_verify_ieee2800 on grid/WindFarm_300MW.pscx. If Low-Voltage Ride-Through (LVRT) reactive current response time $T_{\text{resp}}$ exceeds $30\text{ ms}$, adjust the inverter inner current loop gains and re-run pscad_verify_ieee2800 until all IEEE Std 2800-2022 clauses pass."_

---

Next Steps

• MCP Server & IDE Workbench Reference — Full schema reference for all 52 core MCP tools, IDE copilot tiers, and autonomous social bots.
• CI/CD & GitHub Actions — Automate headless linting, SARIF verification, visual PR diffs, and DHF compliance dossiers on every commit.