SPICEBridge
AI-powered circuit design through simulation — an MCP server that gives language models direct access to SPICE circuit simulation via ngspice, enabling natural language circuit description and automated netlist generation, simulation, measurement, and spec verification.
README
SPICEBridge
<img width="1536" height="1024" alt="image" src="https://github.com/user-attachments/assets/235461c1-122a-41bb-b802-cb6055fb8387" />
AI-powered circuit design through simulation — an MCP server that gives language models direct access to SPICE circuit simulation via ngspice. Describe a circuit in plain English and let the AI handle netlist generation, simulation, measurement, and spec verification.
Features
- 28 tools covering the full circuit design workflow
- 11 built-in templates with automatic component value calculation (E24 series)
- Simulation: AC sweep, transient, DC operating point
- Measurement: bandwidth, gain, DC levels, transient metrics, power
- Monte Carlo & worst-case analysis under component tolerances
- Multi-stage composition — connect circuit stages with automatic port mapping
- Model wizard — generate SPICE
.libmodels from datasheet parameters - KiCad export — output
.kicad_schschematics - Web viewer — interactive schematic viewer in the browser
- Cloud setup wizard — one-command deployment with Cloudflare tunnels (
spicebridge setup-cloud) - Spec verification — compare results against design targets
Install
pip install spicebridge
Requirements
- Python 3.10+
- ngspice installed and on PATH
Quick Start
Local (Claude Code / stdio)
Add to your .mcp.json:
{
"mcpServers": {
"spicebridge": {
"command": "spicebridge"
}
}
}
Cloud (Streamable HTTP)
spicebridge --transport streamable-http --port 8000
Cloud Setup Wizard
One command to go from a local install to a public MCP endpoint:
spicebridge setup-cloud # interactive (named tunnel, permanent URL)
spicebridge setup-cloud --quick # quick tunnel (temporary URL, no account needed)
The wizard handles the full deployment pipeline:
- Installs cloudflared automatically (macOS via Homebrew, Linux via APT) if not found
- Authenticates with Cloudflare (browser-based OAuth, named tunnel mode only)
- Creates or reuses a tunnel — prompts to pick an existing one or make a new one
- Configures DNS routing for your custom domain (named tunnel mode)
- Generates an API key for authentication
- Starts the SPICEBridge server and tunnel together
- Prints connection info with a ready-to-paste JSON config for Claude.ai
Quick tunnels give you a temporary trycloudflare.com URL instantly — no Cloudflare account required. Named tunnels give you a permanent custom domain (e.g. spicebridge.example.com).
Additional options:
spicebridge setup-cloud --domain mcp.example.com # specify custom domain
spicebridge setup-cloud --port 9000 # custom port
spicebridge setup-cloud --no-install # skip cloudflared installation
Example
1. load_template("rc_lowpass_1st", specs={"f_3dB_hz": 1000})
-> netlist with R=1.6k, C=100nF, circuit_id: "a1b2c3d4"
2. run_ac_analysis(circuit_id, start_freq=1, stop_freq=1e6)
-> frequency response data
3. measure_bandwidth(circuit_id)
-> f_3dB_hz: 995
4. compare_specs(circuit_id, specs={"f_3dB_hz": {"target": 1000, "tolerance_pct": 5}})
-> PASS
5. draw_schematic(circuit_id)
-> schematic SVG
Tools
Create & Configure
| Tool | Description |
|---|---|
create_circuit |
Store a SPICE netlist, returns a circuit ID |
delete_circuit |
Delete a stored circuit and clean up resources |
list_templates |
List available circuit templates |
load_template |
Load a template with parameter substitution |
calculate_components |
Calculate component values from target specs |
modify_component |
Change a component value in a stored circuit |
validate_netlist |
Check a netlist for errors before simulation |
Simulate
| Tool | Description |
|---|---|
run_ac_analysis |
AC frequency sweep |
run_transient |
Transient (time-domain) analysis |
run_dc_op |
DC operating point analysis |
Measure
| Tool | Description |
|---|---|
measure_bandwidth |
Find -3 dB bandwidth from AC results |
measure_gain |
Measure gain at a specific frequency |
measure_dc |
Extract DC operating point values |
measure_transient |
Measure time-domain characteristics |
measure_power |
Calculate power dissipation |
Evaluate & Export
| Tool | Description |
|---|---|
get_results |
Retrieve last simulation results |
compare_specs |
Check measurements against target specs |
draw_schematic |
Generate a schematic diagram (PNG/SVG) |
export_kicad |
Export as KiCad 8 schematic (.kicad_sch) |
open_viewer |
Start the interactive web schematic viewer |
Composition & Ports
| Tool | Description |
|---|---|
set_ports |
Define port-to-node mappings for a circuit |
get_ports |
Return port definitions (auto-detect if unset) |
connect_stages |
Compose multiple stages into a single circuit |
Advanced Analysis
| Tool | Description |
|---|---|
run_monte_carlo |
Monte Carlo analysis under component tolerances |
run_worst_case |
Worst-case analysis at tolerance extremes |
auto_design |
Full design loop in one call: template + simulate + verify |
Model Management
| Tool | Description |
|---|---|
create_model |
Generate a SPICE .lib model from datasheet parameters |
list_models |
List all saved models in the model library |
Development
git clone https://github.com/clanker-lover/spicebridge.git
cd spicebridge
pip install -e ".[dev]"
pytest
License
GPL-3.0-or-later
推荐服务器
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。
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
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。