This repository contains the code for our paper "Work Smarter, Not Harder: An LLM-Enabled Agentic Framework for Automated RTL Generation and Debugging".

1. An x86-64 system.
2. Linux/Windows operating system (we used Ubuntu 20.04).
3. Python 3.10 or newer (3.11 is recommended).
4. uv installed and available in PATH.

verilator,vcs, vcf ... should be installed and configured in your system path. (for mcp server)

Install the public SmartCraft agent and its development dependencies:

uv sync --python 3.11 --extra dev

Install your MCP separately:

uv sync --project SmartCraft-MCP --python 3.11 --extra dev

Copy .env.example to .env and configure the OpenAI-compatible model endpoint:

OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxx
OPENAI_API_URL=https://example.com/v1
MODEL_NAME=your-model-name
TIMEOUT=600

Configure the private stdio MCP server in .env. SmartCraft starts it automatically when tools are needed:

SMARTCRAFT_MCP_STDIO=["./SmartCraft-MCP/.venv/bin/smartcraft-mcp"]

SmartCraft first selects an MCP candidate group from the top-level task:

• generate: extract_fsm, solve_kmap, waveform_read, iverilog_verify, and vcs.
• debug: uvllm_solution, meic_solution, assertsolve_model, iverilog_verify, and vcs.

Within the selected group, the LLM reads the tool schemas and dynamically chooses which tools to call from the benchmark contents. waveform_read, uvllm_solution, meic_solution, assertsolve_model, and vcs are currently protocol placeholders: they are discoverable and return a structured not_implemented result, but do not yet run a backend.

The prototype accepts the following input options:

• --task/-t: task kind, either generate or debug.
• --benchmark/-b: path to one benchmark JSON task.

Run the FSM generation task:

uv run smartcraft -t generate -b benchmark/generate/hard/fsm-prob-1-mealy.json

Run the K-map generation task. It uses the same generate task kind; the Agent inspects the MCP tool schemas and dynamically selects the deterministic K-map solver from the benchmark contents:

uv run smartcraft -t generate -b benchmark/generate/hard/kmap-prob-1-null.json

Run the RTL debugging task:

uv run smartcraft -t debug -b benchmark/debug/uvllm/accu_1.json

Generation tasks accept uname, spec or desc, tb, and optional ref fields. Debugging tasks additionally use buggy_code. Each run writes candidate.v and metadata.json to a unique directory under runs/.