OpenEMS electromagnetic simulation MCP server for AI-assisted antenna and RF design
OpenEMS MCP server demonstrates basic antenna design functionality but suffers from significant quality gaps across naming, descriptions, parameter handling, and output documentation. Tool names follow verb_noun convention (design_patch_antenna, design_dipole_antenna, design_horn_antenna), but descriptions are generic and lack LLM-optimized detail. Input schemas are present with type definitions and reasonable defaults, but lack validation guidance. Output schemas are completely undocumented, the tool returns a complex nested dict structure (design object with 'id', 'name', 'type', 'parameters', 'dimensions', 'calculated', 'geometry' fields) but the LLM receives no schema specification of what fields to expect. Error handling is absent: no recovery guidance, no input validation, no categorization of failure modes. The implementation references a complex geometry output structure (_patch_geometry method) but provides no specification of its schema. Parameters accept numeric inputs without documented range constraints or bounds, inviting agents to pass physically unrealistic values (e.g., negative frequencies, implausibly large substrate heights). No security controls are evident, though the tools are read-only and pose minimal risk.
Design a half-wave dipole antenna
Design a pyramidal horn antenna for specified gain
Design a rectangular patch antenna using transmission line model
Output schemas completely undocumented. Tools return complex nested dict structures (design object with dimensions, calculated fields, geometry sub-object) but LLM receives no schema specification. LLMs cannot plan downstream tool chains or extract required fields without documented return types.
Tool descriptions are generic and lack LLM-optimized content. 'Design a rectangular patch antenna using transmission line model' (62 chars) lacks WHEN to use this vs alternatives, WHAT it returns, or practical prerequisites. All three antenna tools have similarly sparse descriptions. Baseline for A+ tools is 50-200 chars with explicit context.
Input parameters lack validation constraints and range documentation. Numeric parameters (frequency_ghz, substrate_er, substrate_height_mm, wire_radius_mm, gain_dbi) accept unbounded float inputs with no documented min/max, physical limits, or engineering constraints. An LLM could pass frequency_ghz=-5 or substrate_er=1e10 without validation guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
No error handling or recovery guidance. Code contains no error classification (retryable vs user-fixable vs fatal), no validation of input ranges, and no actionable error messages. If OpenEMS is unavailable (OPENEMS_IMPORT_ERROR is set), no tool documents this or guides the LLM to check prerequisites.
Parameter descriptions are minimal or missing context. 'Center frequency in GHz' lacks guidance on valid range, typical values, or constraints. 'Wire radius in mm' does not specify minimum/maximum or consequences of invalid input. Engineering domains benefit from explicit bounds: 'frequency_ghz (required, range: 0.1 - 100 GHz; typical for RF/microwave: 0.5 - 60 GHz)'.
No documentation of tool composition or chaining. The three tools are independent antenna designers; it is unclear whether an LLM can use their outputs to compare designs, chain them into a larger workflow, or what downstream tools should accept their output IDs.