Verilog MCP Server

Verilog MCP Server

AI-powered Verilog / SystemVerilog development tools for Claude Code. Generate testbenches, lint RTL code, and scaffold UVM verification environments — all from within your editor.

Status: v0.3.0 — 3 tools, end-to-end verified in Claude Code.

Why

Chip design EDA tools are decades old. AI-assisted hardware development is a blue ocean. This MCP server brings Verilog / SystemVerilog code generation and linting directly into Claude Code, so you can:

• Generate testbench boilerplate from a module declaration in seconds
• Catch common RTL bugs (missing reset, latch inference, case without default) before simulation
• Scaffold a full UVM verification environment with proper clocking blocks, sequencers, and scoreboards

Tools

Tool	Description
generate_testbench	Generate a Verilog testbench skeleton from a module definition
check_lint	Static lint checks: missing reset, case-default, latch risk, width mismatch
generate_uvm_env	Generate a 10-component UVM verification environment

generate_testbench

Input: a Verilog module. Output: a complete testbench with clock generation, reset sequence, DUT instantiation, $dumpfile / $dumpvars, and a TODO placeholder for custom test stimuli.

• Auto-detects clock and reset ports
• Handles active-low reset (rst_n suffix)
• Supports both new-style (input clk, input rst_n) and old-style (input clk, rst_n) port declarations
• 100MHz default clock (10ns period)

check_lint

Four categories of checks:

1. Sequential always block without reset — WARNING
2. case/casez/casex without default — ISSUE (latches / X-states)
3. Incomplete if-assignments in combinational always blocks — WARNING (latch inference risk)
4. Width mismatch heuristics — INFO

generate_uvm_env

Generates 10 UVM component files separated by // ===== FILE: filename.sv ===== markers:

interface → transaction → sequence → driver → monitor → agent → scoreboard → environment → test → top

• Proper drv_cb (outputs-only clocking block) and mon_cb (inputs-only) design
• uvm_analysis_port for monitor → scoreboard communication
• Active/passive agent mode
• TODO placeholders in scoreboard and constraints

Installation

# 1. Clone
git clone https://github.com/kaiwenyu147369/verilog-mcp-server.git
cd verilog-mcp-server

# 2. Create virtual environment
python -m venv venv

# 3. Activate & install
# Windows:
venv\Scripts\activate
# macOS / Linux:
source venv/bin/activate

pip install mcp

Usage with Claude Code

Add to your workspace .mcp.json:

{
  "mcpServers": {
    "verilog-mcp-server": {
      "command": "C:\\path\\to\\venv\\Scripts\\python.exe",
      "args": ["C:\\path\\to\\verilog-mcp-server\\server.py"]
    }
  }
}

Then in Claude Code settings (~/.claude/settings.json):

{
  "enableAllProjectMcpServers": true,
  "enabledMcpjsonServers": ["verilog-mcp-server"]
}

Restart VSCode, and the tools will appear in Claude Code.

Quick Test

# Activate venv first, then:
python test_uvm.py
python test_v2.py

Tech Stack

• Python MCP SDK 1.27.1 — @server.list_tools() + @server.call_tool() (two-decorator API)
• stdio JSON-RPC — Claude Code communicates via stdin/stdout
• Regex-based parsing — No SystemVerilog compiler dependency

Known Limitations

• Port parser uses regex, not a full SystemVerilog tokenizer — deeply nested parentheses in complex parameterized ports may not parse correctly
• UVM Scoreboard comparison logic is a TODO placeholder
• Lint checks are heuristic; no full elaboration or type system
• No interface / struct / modport support in port parsing

Roadmap

• v0.4.0 — 4th tool (generate_assertions or analyze_waveform)
• Port parser upgrade to simple tokenizer
• Additional lint rules (blocking/non-blocking assignment check, signed/unsigned warnings)
• UVM Scoreboard auto-comparison logic
• Multi-file project support

License

MIT