FilterForge: An LLM-Based, Semi-Automated Agentic VS Code Extension for Microwave Bandpass Filter Design

We present FilterForge, a chat-driven VS Code environment that pulls the synthesis, analysis, simulation, and optimization stages of microwave bandpass filter design, normally coordinated by hand across tools written in different languages, into one workflow. A deployed Model Context Protocol (MCP) server exposes deterministic Python implementations of coupling-matrix synthesis, uniform predistortion, topology reconfiguration, a genetic-algorithm transmission-zero selector, a mode-matching engine for H-plane iris-coupled rectangular waveguide geometries, and a skill that generates PyAEDT/HFSS notebooks for various dimensioning design-curves. A language-model orchestrator turns natural-language requests into typed tool calls, while every reported quantity stays inside the deterministic kernels, so the numerics remain reproducible and model-agnostic. We evaluate the call layer on a 45-task benchmark across the five tool categories: gemini-3-flash reaches (Formula presented.) tool-selection and (Formula presented.) full-call accuracy with an (Formula presented.) pass3 rate, which an ablation traces to the curated tool-selection prompt rather than to raw model capability. The mode-matching engine is validated against full-wave HFSS on a six-pole 4 GHz Chebyshev filter tuned from the chat panel, and on an 8 GHz WR-112 counterpart taken end-to-end with no engineer in the loop, where a deterministic critique gates each round until a manufacturable geometry is reached. We then exercise the full workflow on two folded six-pole WR-90 cross-coupled filters at (Formula presented.), a high-selectivity design synthesized against a stop-band mask and a group-delay-equalized variant whose positive cross-coupling uses a pair of side-wall irises, the latter settling to a peak-to-peak in-band group-delay ripple below (Formula presented.) while recovering the synthesized return loss. © 2026 by the authors.