Show HN: SPICE simulation → oscilloscope → verification with Claude Code
TL;DR Highlight
This is an experimental case demonstrating that connecting a SPICE simulator and a real oscilloscope to Claude Code via an MCP server allows for creating a feedback loop where AI directly analyzes and verifies simulation results and actual waveform data.
Who Should Read
Firmware/hardware engineers who want to utilize LLMs in hardware design or embedded development, or developers curious about cases of connecting MCP servers to actual physical equipment.
Core Mechanics
- Simply 'describing circuit designs in natural language and having Claude create the circuit' works for simple circuits but quickly reaches its limits in complex designs. This is because the AI often fails to understand what is desired accurately.
- However, it was much more effective to adopt a method of connecting Claude Code to immediate feedback means (oscilloscope, SPICE simulator) as tools. Claude shines in an environment where feedback comes immediately.
- The biggest benefit in this workflow was data analysis. Claude automatically handles tedious tasks that were previously done visually, such as time axis normalization and data sorting.
- Two MCP servers were created: `lecroy-mcp` for controlling a LeCroy oscilloscope and `spicelib-mcp` wrapping a SPICE simulator library. This allows Claude to call the actual measurement equipment and simulator as tools.
- Important principle when connecting an oscilloscope: Since Claude cannot see the physical connection status, never let it guess what is connected where. You must manage to prevent old measurement data from remaining in the context.
- Do not paste raw measurement data directly into Claude's context, but save it to a file and have Claude indirectly read and process the file. This is a practical tip to prevent context contamination.
- When working with microcontrollers, explicitly provide the pinout/pinmux map and prepare a Makefile exposing functions like build/flash/ping/erase in advance to prevent Claude from improvising commands.
- This demo is configured with intentionally simple circuits like an RC filter, but the author explains that this approach scales well to much more complex circuits and real embedded projects.
Evidence
- "There were experiences trying to design boards with Claude Opus and SPICE simulations, but the board functionality was completely hallucinated, and it made absurd claims like 'discovering a secret $1 billion project'. Ultimately, none of the boards worked, and it was busy praising itself without realizing it had created gibberish.\n\nSimilarly, there was an opinion that attempting to automate circuit design/simulation/optimization was quite good at reading and explaining files, but could not properly modify designs beyond the simplest circuits. This is because it does not understand the physical concepts behind the circuit, and ultimately the agent was properly utilized only after defining the range of config files (step size, optimization algorithm, min/max, etc.) to be passed to the simulator in a script.\n\nA case of trying to automate KiCad schematic review was also shared, and the same wall was hit. Initially, having Claude directly grep/read the `.kicad_sch` file resulted in creating non-existent pin numbers, but the problem was mostly solved by generating JSON with a Python analyzer and having Claude read only the JSON. There were also technical questions about how `spicelib-mcp` handles models not in the basic library (whether it passes the `.lib` path as a tool arg or whether the server manages the registry directly). \n\nA practical tip was shared in the comments that using Mermaid diagrams makes it much less cumbersome to explain and modify circuits to LLMs.\n\nThere was also an opinion that local LLMs should be used instead of commercial services like Claude. The argument was that local LLMs are the future, even if they are slower, pointing out the frequent policy changes (rug-pull), price increases, and user-unfriendly direction of commercial services."
How to Apply
- If you have a lot of manual work comparing SPICE simulation results with actual oscilloscope measurements, you can connect `lecroy-mcp` and `spicelib-mcp` to Claude Code to automate waveform data analysis and alignment. This can significantly reduce the burden of repetitive preprocessing tasks like time axis normalization.
- To automate embedded firmware development with Claude Code, prepare a Makefile containing build/flash/ping/erase actions in advance and explicitly provide the pinout map to prevent Claude from creating commands on the fly and causing malfunctions.
- Pasting raw data from measurement equipment directly into Claude's context wastes tokens and contaminates the context. Instead, save the data to a file and configure Claude to access it indirectly via the file path to create a more stable analysis workflow.
- When connecting physical equipment to MCP, do not let Claude guess the physical connection status, but explicitly tell it what is connected to which channel at the start of the session. Also, periodically reset the session state to prevent previous measurement data from remaining in the context to obtain reliable results.
Terminology
Related Papers
ChatGPT for Google Sheets Exfiltrates Workbooks
Google Sheets용 ChatGPT 확장 프로그램이 간접 프롬프트 인젝션 공격에 취약해, 단 하나의 시트에 숨겨진 악성 명령만으로 계정 내 워크북 전체가 외부로 유출될 수 있다는 보안 연구 결과가 공개됐다.
LinTree: Improving LLM Reasoning with Explicitly Structured Search Histories
LLM의 추론 트레이스에 부모 포인터(parent pointer)만 추가해도 탐색 성능과 효율이 크게 올라간다.
Diagnosing Failure Modes of Shared-State Collaboration in Resource-Constrained Visual Agents
4B~8B 소형 비전 모델에서 공유 메모리(화이트보드) 기반 멀티에이전트 협업이 오히려 성능을 떨어뜨리는 이유를 분석한 연구.
Show HN: Open Envelope – an open schema for defining AI agent teams
여러 AI 에이전트가 팀처럼 협업하는 구조를 벤더 중립적인 오픈 스키마로 선언적으로 정의할 수 있게 해주는 프로젝트로, 멀티 에이전트 오케스트레이션의 표준화를 시도한다.
Unifying Temporal and Structural Credit Assignment in LLM-Based Multi-Agent Prompt Optimization
여러 AI 에이전트가 협력할 때 '어느 라운드의 어느 에이전트'가 실패했는지 정확히 짚어내서 그 프롬프트만 고치는 최적화 프레임워크
Show HN: Ktx – Open-source executable context layer for data agents
AI 에이전트가 회사 데이터 웨어하우스를 정확하게 쿼리할 수 있도록 시맨틱 레이어, 메모리, 비즈니스 지식을 자동으로 구축해주는 오픈소스 도구다. 기존 에이전트가 매번 웨어하우스를 재탐색하거나 잘못된 메트릭 로직을 임의로 만들어내는 문제를 해결한다.
Related Resources
- Original blog post
- lecroy-mcp: MCP server for LeCroy oscilloscopes (GitHub)
- spicelib-mcp: SPICE simulator wrapper MCP server (GitHub)
- rc-filter-demo-files: Demo setup files (GitHub)
- kicad-happy: KiCad schematic review automation (mentioned in comments)
- SPICE Wikipedia
- FPAA related EE Times podcast (mentioned in comments)