AI-powered circuit design bridge connecting language models to ngspice simulation via MCP
SPICEBridge provides 27 tools with generally clear naming and decent descriptions, but exhibits significant gaps in schema completeness, parameter documentation, and error handling guidance. Naming follows verb_noun convention (create_, get_, list_, delete_, run_, draw_, modify_, validate_, parse_, format_, snap_, compose_, export_, generate_, read_, solve_, auto_). Descriptions are present for all tools and mostly adequate (ranging 40-200 chars), but many lack actionable recovery guidance. Parameter schemas are visible and typed, but descriptions are often generic or missing details about constraints, ranges, and valid values. Error handling is minimal, most tools return error dicts without recovery guidance. Output schemas are not documented. The server demonstrates competent implementation of core SPICE features but falls short of production-grade tool design standards.
Automatically generate circuit design from description and templates
Auto-detect port roles from netlist node names using heuristics
Compose multi-stage circuit from templates with port connections
Create a new circuit with a given SPICE netlist
Delete a circuit and clean up its associated files
Generate a schematic diagram from a netlist
Export circuit to KiCad 8 schematic format
Missing output schema documentation for all 27 tools. LLMs cannot plan subsequent calls or extract required fields without knowing return structure. E.g., create_circuit likely returns a circuit_id, but this is not documented.
Insufficient error handling guidance. Error responses return simple dicts like {"status": "error", "error": "..."}. No recovery guidance, what should the LLM do next? E.g., run_simulation fails with 'simulation timed out', should it retry, adjust parameters, or fall back to a different tool?
Parameter descriptions lack constraint details. E.g., num_runs in monte_carlo_analysis is described as 'Number of simulation runs' with no mention of the MAX_MONTE_CARLO_RUNS=100 limit visible in server.py. LLMs may request values that exceed limits.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 54 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 48 | - | v1 |
Format a numeric value using engineering notation (k, M, m, u, n, p)
Generate a SPICE model from behavioral specification
Retrieve the stored netlist and metadata for a circuit
Retrieve a specific circuit template
List all stored circuits with their IDs and result status
List available SPICE component models
List available circuit templates
Modify component values in a netlist
Run Monte Carlo simulation with component value randomization
Parse SPICE netlist and return structured component information
Parse ngspice output results file into structured data
Parse a SPICE value string (e.g., 10k, 2.2u) into a numeric value
Extract AC analysis results at a specific frequency
Calculate bandwidth from AC analysis results
Run ngspice simulation on a circuit and return results
Compute sensitivity of output to component value changes
Round component values to standard series (E6, E12, E24, etc.)
Solve for component values to meet design specifications
Validate SPICE netlist syntax without running simulation
Perform worst-case corner analysis on circuit
auto_design tool combines multiple concerns (generation + composition). Description says 'Automatically generate circuit design from description and templates' but parameters include 'ports' (mapping), suggesting composition logic is entangled. Should split into auto_design + compose_stages.
Parameter naming inconsistency. Tools use 'circuit_id' but the description in compose_stages uses 'template_name' and 'label' without suffix hints (template_name vs template, is 'label' a string or ID?). Mixed conventions confuse LLMs.
No pagination or limit guidance for list_* tools. list_circuits, list_templates, list_models have no documented max results or pagination fields. If hundreds of templates exist, responses could blow context windows.
Destructive operations (delete_circuit, modify_netlist, export_kicad) lack confirmation/dry-run patterns. No dry-run parameter to preview changes. delete_circuit irreversibly removes circuit data, agents may trigger this by mistake.
Tool response includes internal schematic_url fields used for display, but tool definitions do not document this output field. LLMs cannot extract and use these URLs without guidance.