schematic-mcp-bridge

schematic-mcp-bridge

Enables AI agents to design electrical schematics through a high-level semantic API, handling components, pins, nets, validation/ERC, auto-layout, and SVG/KiCad rendering without requiring coordinate or file internals.

Category
访问服务器

README

schematic-mcp

AI-agent-friendly system for building electrical schematics: an MCP server exposes a high-level, semantic API (components, pins, nets) so an LLM never has to touch coordinates, wires, or .kicad_sch internals directly.

AI Agent → MCP / high-level API → semantic schematic model → validation/ERC
    → automatic layout → SVG and/or KiCad schematic

See schematic_mcp_requirements.md for the full spec and SKILL.md for agent rules.

Status

  • Stage 1 done: minimal semantic model (Component, Pin, Net, Schematic) with JSON save/load, in schematic/model.py.

  • Stage 2 done: KiCad symbol library import backed by SKiDL (schematic/library.py: search_components / get_component_pins / instantiate) and structural validation + basic ERC (schematic/validation.py: validate / run_erc). Tested against real official KiCad symbol libraries vendored in tests/fixtures/kicad-symbols/ (Device, MCU_Espressif, Sensor_Current).

  • Stage 3 done: naive auto layout (schematic/layout.py, one row of boxes, no overlap), an SVG renderer (schematic/renderer.py) that draws component boxes with pin numbers/names and net-name labels next to each pin — no drawn wires yet, connectivity is shown via labels (see requirements section 11); and a CLI (cli/main.py: search, validate, render).

  • Stage 4 done: MCP server (mcp_server/server.py) exposing the full section 8 tool list — library lookup, schematic lifecycle, connections, layout hints, validation/ERC, SVG output — over stdio. The stateful logic lives in mcp_server/tools.py::SchematicSession (one active schematic per session), framework-free so it's tested directly without an MCP client. set_placement_hint / set_pin_side / group_components store intent on the model already.

  • Stage 5 done: auto_layout now honors left_of/right_of placement hints (topological reorder of the row; other relations are still ignored, a real 2D layout is stage 7 territory if needed). New schematic/routing.py draws an actual orthogonal wire between two components when they land as immediate neighbors and it's a 2-endpoint net with pins facing each other — otherwise still a net label, since routing around an intervening box is real routing-algorithm work (deferred). set_pin_side/group_components are still stored-but-unconsumed.

  • Stage 6 done: .kicad_sch export (schematic/exporters/kicad.py), reusing SKiDL's own schematic generator rather than hand-rolling a KiCad file writer — it already draws real classical symbols with correct geometry (section 4/19's stated preference). Only works for components sourced from a real library symbol via add_component; raises KicadExportError on a hand-authored/invented component or one that doesn't resolve, rather than silently producing a broken file. examples/esp32_ina226.json can't export (its ESP32/INA226 pins are hand-authored placeholders, not real symbols) — see examples/build_rc_filter.py for a small schematic built entirely from real library parts that does. Needed vendoring power.kicad_sym too (SKiDL's schematic generator loads it unconditionally).

    Also fixed a real bug found while building that RC filter example: ComponentLibrary.instantiate() defaulted reference to the bare library ref_prefix ('R' for every resistor) instead of component_id — so adding a second resistor without an explicit reference= produced a false duplicate_reference ERC error on a perfectly valid schematic.

MVP is now complete end-to-end (search → add → connect → validate/ERC → SVG / KiCad export) via both the CLI and the MCP server. What's left is optional polish: stage 7 (ELK-based 2D layout + general wire routing) is explicitly "only if this isn't good enough" territory in the spec, not committed to.

Development

python3 -m pip install --user --break-system-packages -e ".[dev]"
PYTHONPATH=. python3 -m pytest

# CLI
PYTHONPATH=. python3 -m cli.main search resistor
PYTHONPATH=. python3 -m cli.main validate examples/esp32_ina226.json
PYTHONPATH=. python3 -m cli.main render examples/esp32_ina226.json -o out.svg
PYTHONPATH=. python3 -m cli.main export-kicad examples/rc_filter.json -o out.kicad_sch

# MCP server (stdio)
PYTHONPATH=. python3 -m mcp_server.server

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选