Silvaco Handbook MCP

中文文档

An MCP (Model Context Protocol) server that gives AI assistants fast, bounded access to the Silvaco TCAD manuals and the official Deckbuild example decks — without dumping thousands of PDF pages into the conversation context.

It indexes MinerU-converted Markdown of the manuals into a local SQLite FTS5 database and exposes context-friendly search/read tools over stdio. Each manual section carries its original PDF page range, so results can be cross-referenced with the source PDFs.

Features

• Handbook corpus — section-level index of the Silvaco manuals (Deckbuild, Victory Device, Victory Process, ...) with original PDF page ranges preserved.
• Examples corpus — full-text index of the official .in example decks from your Silvaco installation (examples/deckbuild/<version>).
• Bounded output — every tool has hard caps (search ≤ 30 hits, read ≤ 20 000 chars, paged via offset) so the LLM context stays small.
• Incremental indexing — the index is rebuilt lazily on startup whenever a Markdown file or example deck changes (size/mtime fingerprint).

Tools

Tool	Description
handbook_list_manuals	List indexed manuals with section counts and source paths
handbook_toc	Section headings of a manual (its table of contents), optional title filter
handbook_search	Full-text search; returns section id, title, PDF page range, highlighted snippet
handbook_read	Read one section by id; page through long sections with offset
examples_list	List indexed example decks (name / category / description)
examples_search	Full-text search over deck name, description, and content
examples_read	Read one example deck in full (partial unique names accepted)

Requirements

• Python ≥ 3.10
• Python packages: mcp, pymupdf (see requirements.txt)
• MinerU Open API CLI (mineru-open-api) plus an API token — only needed for convert_mineru.py
• The Silvaco manuals as PDFs, and (optionally) a Silvaco installation for the example decks

Install the Python dependencies:

pip install -r requirements.txt

Quick start

1. Convert the manuals to Markdown

MinerU's precision extract is limited to 200 pages per request, so convert_mineru.py processes each PDF in page-range chunks, marks chunk boundaries with <!-- pdf pages S-E --> comments, and merges them into one Markdown file per manual. It is resumable: existing chunks are skipped on re-run.

export MINERU_TOKEN=<your-mineru-token>
export SILVACO_PDF_DIR=/path/to/silvaco/handbook/pdfs
export SILVACO_MD_DIR=/path/to/markdown/output

# Convert every PDF in SILVACO_PDF_DIR (or only one manual with --only)
python convert_mineru.py
python convert_mineru.py --only deckbuild_users1 --chunk-size 200

Output layout:

<SILVACO_MD_DIR>/<manual>/chunks/p001-200.md   # one per chunk
<SILVACO_MD_DIR>/<manual>/chunks/images/      # extracted images
<SILVACO_MD_DIR>/<manual>/<manual>.md         # merged, indexed by the server

2. Build the index

python server.py --build          # incremental
python server.py --build --force  # full rebuild

The index also rebuilds itself lazily when the server starts and a source file has changed, so this step is optional.

3. Register the MCP server

Add it to your MCP client configuration.

Kimi Code / Claude Code (.kimi-code/mcp.json or .claude/mcp.json):

{
  "mcpServers": {
    "silvaco-handbook": {
      "command": "python",
      "args": ["/path/to/Silvaco_MCP/server.py"],
      "env": {
        "SILVACO_MD_DIR": "/path/to/markdown/output",
        "SILVACO_EXAMPLES_DIR": "C:/Silvaco/examples/deckbuild/5.2.40.R"
      }
    }
  }
}

Claude Desktop (claude_desktop_config.json): same mcpServers block.

Cursor: Settings → MCP → add the same command/args/env.

4. Ask questions

Once connected, your assistant can answer with manual citations, e.g.:

• "How does Atlas model impact ionization with the Selberherr model?" → handbook_search("impact ionization Selberherr") → handbook_read(...)
• "Show me an official example of a quantum well laser gain simulation." → examples_search("quantum well laser optical gain") → examples_read(...)

Configuration

All paths are set via environment variables:

Variable	Default	Purpose
SILVACO_MD_DIR	<repo>/../../mineru-output/silvaco	Root of converted manual Markdown (<manual>/<manual>.md)
SILVACO_CACHE_DB	<repo>/.cache/handbooks_md.db	SQLite FTS5 cache location
SILVACO_EXAMPLES_DIR	(unset → examples tools disabled)	Root of the official deckbuild examples tree
SILVACO_PDF_DIR	(vault layout default)	Handbook PDF directory (used by convert_mineru.py)
SILVACO_MCP_CONFIG	(vault layout default)	mcp.json to read MINERU_TOKEN from (optional)
MINERU_TOKEN	—	MinerU Open API token (takes precedence over config file)
MINERU_CLI	mineru-open-api	Path/name of the MinerU CLI executable

Usage examples

Typical agent workflow with the tools:

handbook_list_manuals()
→ [{"manual": "deckbuild_users1", "sections": 481, ...}, ...]

handbook_search("impact ionization Selberherr", manual="victorydevice")
→ [{"section_id": 512, "section": "3.7.4 Impact Ionization Models",
    "pdf_pages": "201-400", "snippet": "... **Selberherr** ..."}, ...]

handbook_read("victorydevice", section_id=512)
→ "===== victorydevice [512] 3.7.4 Impact Ionization Models (pdf p.201-400) =====
   ... full section text ..."

handbook_read("victorydevice", section_id=512, offset=12000)   # continue long sections

examples_search("quantum well laser optical gain", category="Opto")
→ [{"name": "optoex14", "category": "Technology/Opto_and_Photonics",
    "description": "Quantum Well Laser ...", "snippet": "..."}, ...]

examples_read("optoex14")
→ "===== example optoex14 [...] =====\n# Quantum Well Laser ...\ngo atlas ..."

Maintenance

• Manual PDFs updated → re-run python convert_mineru.py (incremental); the index rebuilds automatically on the next server start, or force it with python server.py --build.
• Example decks changed under SILVACO_EXAMPLES_DIR → the index is updated incrementally on startup (deleted decks are dropped as well).
• The .cache/ directory is derived state — safe to delete at any time.