CadenceAI MCP — Cadence EDA Tools for AI Assistants

A Model Context Protocol (MCP) server that exposes Cadence EDA tools (OrCAD Capture, Allegro PCB Designer, PSpice) to AI assistants like Claude, Cursor, and Cline — letting them browse PCBs, generate schematics, and run circuit simulations through a single, license-free, cross-platform interface.

✨ Features

13 MCP tools, 75+ Allegro batch subcommands wrapped with safety allow-list, 6 circuit templates, all available out of the box.

Related MCP server: KiCAD MCP Server

🚀 Quick Start

1. Install

git clone https://github.com/cadenceai/cadence-mcp.git
cd cadence-mcp
pip install -e .

2. Probe your environment

python -m cadence_mcp.server --probe

{
  "cds_root": "C:\\Cadence\\SPB_23.1",
  "allegro_batch": "C:\\Cadence\\SPB_23.1\\tools\\bin\\allegro_batch.exe",
  "pspice_exe": "C:\\Cadence\\SPB_23.1\\tools\\bin\\pspice.exe",
  "license_present": false,
  "batch_commands_count": 75,
  ...
}

3. Register in your MCP client (e.g. Cursor / Claude Desktop)

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "cadence": {
      "command": "python",
      "args": ["-m", "cadence_mcp.server"],
      "cwd": "D:\\AIBench\\CadenceAI"
    }
  }
}

Restart the client. AI now has 13 Cadence tools at its disposal.

🛠 The 13 Tools

Environment & Diagnostics

PCB Design

Schematic & Netlist

Simulation

📐 Architecture

                  ┌──────────────────────────┐
   AI Client ───► │  MCP Server (stdio)      │
  (Claude /       │   13 tools               │
   Cursor)        │                          │
                  └────────────┬─────────────┘
                               │
       ┌───────────────────────┼─────────────────────────┐
       │                       │                         │
       ▼                       ▼                         ▼
 ┌──────────────┐      ┌──────────────────┐      ┌──────────────┐
 │ env_detect   │      │ simulator        │      │ schematic    │
 │ pcb.py       │      │   (PySpice/PSpice)│      │ pcb.py       │
 │  (read-only) │      │                  │      │  (text gen)  │
 └──────────────┘      └──────────────────┘      └──────────────┘
       │                       │                         │
       ▼                       ▼                         ▼
   Allegro .brd            PSpice / ngspice           Allegro netin
   files / dirs            binary backend             physical netlist

Three independent modules + one MCP server bootstrap. No monolithic controller, no hidden state beyond an env cache.

🧪 Testing

python tests/test_all.py

Covers 10 scenarios end-to-end:

TEST  1: 环境探测                [OK] CDSROOT 已识别, 75+ batch commands
TEST  2: Allegro .brd 扫描       [OK] 找到 N 个 .brd 文件
TEST  3: SPICE 网表生成          [OK] .TRAN / .END / 文件已保存
TEST  4: 电路 DSL 解析           [OK] 识别元件、控制指令
TEST  5: 生成 OrCAD DSN 文本     [OK] XML 写入成功
TEST  6: SPICE → Allegro 转换    [OK] 包含 R1 / C1 等 refdes
TEST  7: 安全命令清单            [OK] 56+ 命令白名单
TEST  8: allegro_batch help 调用 [OK] exit=0
TEST  9: MCP 工具注册表          [OK] 13 tools / 13 handlers
TEST 10: 端到端电路设计流程      [OK] .cir / .tel / .dsn 全部产出

Live MCP demo

For a real MCP-protocol walkthrough (spawns the server, calls every tool via JSON-RPC stdio, shows actual outputs and writes real files):

python demo_mcp_live.py

Expected outcome: all 14 tasks succeed; produces _demo_rc.cir, _demo_div.cir, _demo_step.cir, _demo_rc.tel, _demo_amp.dsn in examples/.

For just the simulation numeric results without MCP overhead:

python demo_summary.py

🧩 End-to-End Example

Tell your AI assistant:

"Design a low-pass RC filter with fc=160 Hz and simulate it."

The AI calls, in order:

1. cadence_probe                       # learn environment
2. list_simulation_templates           # see what's available
3. generate_spice_netlist              # template=rc_lowpass, fc≈160Hz
4. run_circuit_simulation              # execute on PSpice/ngspice
5. netlist_to_allegro                  # convert for PCB layout
6. allegro_batch_run subcommand=netin  # import into Allegro

Output flow:

ai_design.cir    PSpice/ngspice  →  ai_design.out (measurements)
ai_design.tel    SPICE → Allegro  →  ready for `allegro_batch netin`
ai_design.dsn    OrCAD Capture   →  optional: hand-edit in Capture

🔧 Simulation Backends

The server auto-selects the best available backend in this order:

Auto-skip behavior: if cadence_probe reports license_present: false, the server skips PSpice automatically (avoids the 60s timeout) and goes straight to the analytical fallback. So you get instant simulation results even without any license.

Tested backends:

• PSpice binary present at C:\Cadence\SPB_23.1\tools\bin\pspice.exe; -b mode requires an active license. When license is missing, server falls back gracefully.

• ngspice is the recommended open-source fallback. Download from https://ngspice.sourceforge.io/download.html and put ngspice.exe on PATH. (On some networks SourceForge blocks direct downloads; try MSYS2 mingw-w64-ucrt-x86_64-ngspice instead.)

• PySpice 1.5 Python wrapper is auto-detected (pyspice_available: true).

• Analytical fallback (numpy only) — solves RC low/high-pass, step response, and voltage dividers in closed form / closed-form-plus-numerical. Always available, no external binaries required.

Demonstrated simulation results (no license required)

Run python demo_summary.py:

RC 低通滤波器 (rc_lowpass)     Vout_peak=4.36V   tau=100μs   fc=1.59kHz
RC 高通滤波器 (rc_highpass)    Vout_peak=4.36V   tau=100μs   fc=1.59kHz
RC 阶跃响应  (rc_step)         Vout_end=0.48V    tau=10ms    fc=15.9Hz
电压分频器   (voltage_divider) Vout=8.0V  (Vin=12V, R1=1k, R2=2k)

🛡 Security & Sandboxing

• All external commands run through subprocess.run with hard timeout.

• allegro_batch_run enforces a 56-command allow-list unless allow_unsafe=true is explicitly set.

• File I/O paths are normalised; netlist files are written with ASCII encoding to avoid encoding edge cases.

• No network calls; pure local execution.

• License state is reported (license_present: false/true) but never assumed.

📚 Reference: Useful Allegro Batch Subcommands

Selected from the 75+ available:

Run allegro_batch.exe help on your machine to get the full list.

🤝 Contributing

PRs should:

1. Run python tests/test_all.py cleanly (no failures).

2. Add tests for any new tool.

3. Update this README's tool table.

4. Keep the allow-list curated — every batch command added must be safe.

📄 License

MIT — see LICENSE.

🔗 Related Projects

• allegro-mcp — read-only MCP for .brd files (MIT, 4 stars).

• skill-script-generator — NL → SKILL script generator.

• Python-Skill Bridge — Python ↔ Virtuoso SKILL bridge.

• mcp4eda.cn — Chinese EDA MCP hub (no Cadence entries yet).

CadenceAI MCP is the first combined schematic + PCB + simulation MCP for the Cadence SPB toolchain.