@claymore-dev/anvil

MCP server that makes your project docs searchable by AI. Point it at a folder of markdown files, and it chunks, embeds, and serves them over the Model Context Protocol — so your AI assistant can actually find relevant documentation.

Quickstart

npm install -g @claymore-dev/anvil
cd your-project
anvil serve --docs ./docs

That's it. Anvil indexes your markdown, starts an MCP server on stdio, and exposes 5 tools your AI can call.

What It Does

1. Scans a directory for .md files
2. Chunks them by heading structure (not fixed-size blocks)
3. Embeds each chunk using a local transformer model (no API keys needed)
4. Stores chunks + vectors in a local SQLite database with sqlite-vss
5. Watches for file changes and re-indexes automatically
6. Serves semantic search + page/section retrieval over MCP (stdio transport)

Installation

npm install -g @claymore-dev/anvil

Requires Node.js 18+. The embedding model (~80 MB) downloads automatically on first run.

Configuration

Create anvil.config.json in your project root (or run anvil init):

{
  "docs": "./docs",
  "db": "./.anvil/index.db",
  "embedding": {
    "provider": "local",
    "model": "all-MiniLM-L6-v2"
  },
  "chunking": {
    "maxChunkSize": 6000,
    "minChunkSize": 200,
    "mergeShort": true
  },
  "watch": true,
  "logLevel": "info"
}

All fields are optional — defaults are shown above.

| Field | Description | |---|---| | docs | Path to markdown directory (default: ./) | | db | SQLite database path (default: ./.anvil/index.db) | | embedding.provider | "local" (default) or "openai" | | embedding.model | HuggingFace model name (default: all-MiniLM-L6-v2) | | chunking.maxChunkSize | Max characters per chunk before splitting (default: 6000) | | chunking.minChunkSize | Short chunks below this merge into parent (default: 200) | | chunking.mergeShort | Enable short-chunk merging (default: true) | | watch | Watch for file changes (default: true) | | logLevel | silent, error, warn, info, or debug |

CLI flags override config file values. Config file overrides defaults.

MCP Client Setup

Cursor

In .cursor/mcp.json:

{
  "mcpServers": {
    "anvil": {
      "command": "anvil",
      "args": ["serve", "--docs", "./docs"]
    }
  }
}

Claude Desktop

In ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "anvil": {
      "command": "anvil",
      "args": ["serve", "--docs", "/absolute/path/to/your/docs"]
    }
  }
}

OpenClaw / mcporter

{
  "mcpServers": {
    "anvil": {
      "command": "anvil",
      "args": ["serve", "--docs", "/absolute/path/to/your/docs"]
    }
  }
}

CLI Reference

anvil serve (default)

Start the MCP server over stdio.

anvil serve [options]

| Flag | Description | |---|---| | -d, --docs <path> | Docs directory (default: ./) | | --db <path> | Database path | | -c, --config <path> | Config file path (default: ./anvil.config.json) | | --no-config | Ignore config file | | --no-watch | Disable file watcher | | --max-chunk-size <n> | Max chunk size in characters | | --min-chunk-size <n> | Min chunk size before merge | | --embedding-provider <p> | local or openai | | --log-level <level> | silent, error, warn, info, debug | | -v, --version | Print version |

anvil index

One-shot indexing without starting the MCP server.

anvil index [options]

Same flags as serve, plus:

| Flag | Description | |---|---| | --force | Clear index and re-embed everything |

anvil init

Create anvil.config.json interactively.

anvil init [--yes]

| Flag | Description | |---|---| | --yes | Skip prompts, use all defaults |

MCP Tools

Anvil exposes 5 tools to connected MCP clients:

| Tool | Description | |---|---| | search_docs | Semantic search across all indexed docs | | get_page | Retrieve full page content by file path | | get_section | Retrieve a specific section by heading path | | list_pages | List all indexed pages with metadata | | get_status | Server health, index state, and version info |

Architecture Overview

┌─────────────┐     ┌──────────┐     ┌──────────┐
│  .md files  │────▶│ Chunker  │────▶│ Embedder │
└─────────────┘     └──────────┘     └──────────┘
                         │                 │
                         ▼                 ▼
                    ┌─────────────────────────┐
                    │   SQLite + sqlite-vss   │
                    └─────────────────────────┘
                              │
                              ▼
                    ┌─────────────────────┐
                    │  MCP Server (stdio) │
                    └─────────────────────┘
                              │
                    ┌─────────────────────┐
                    │  AI Client (Cursor, │
                    │  Claude Desktop...) │
                    └─────────────────────┘

• Chunker (src/chunker.ts) — Splits markdown by heading structure using remark AST. Handles frontmatter, duplicate headings, short-section merging, and long-section splitting.
• Embedder (src/embedder.ts) — Generates 384-dim vectors using all-MiniLM-L6-v2 via @huggingface/transformers. Runs entirely local, no API keys.
• Database (src/db.ts) — SQLite with sqlite-vss for vector similarity search. WAL mode for concurrent reads.
• Indexer (src/indexer.ts) — Diffing engine that only re-embeds changed chunks. Detects model changes and triggers full re-embed.
• Watcher (src/watcher.ts) — Chokidar-based file watcher with debouncing. Re-indexes on add/change/delete.
• Query Engine (src/query.ts) — Vector search, page retrieval, section retrieval, and page listing.
• Server (src/server.ts) — Orchestrates startup: scan → index → watch → MCP.
• Tools (src/tools/index.ts) — MCP tool definitions and handlers.

Contributing

1. Fork the repo
2. Create a feature branch: git checkout -b my-feature
3. Run tests: npm test
4. Submit a PR

License

MIT © Dan Hannah