OpenCode LLM Wiki - MCP Server

MCP (Model Context Protocol) server for OpenCode LLM Wiki - Knowledge graph backend with vector search.

📦 Installation

From npm

npm install -g @opencode-llm-wiki/mcp-server

From source

git clone https://github.com/Moowningstar/opencode-llm-wiki.git
cd opencode-llm-wiki/src-mcp
npm install

🚀 Quick Start

Usage in Claude Desktop

Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "llm-wiki": {
      "command": "node",
      "args": [
        "/path/to/node_modules/@opencode-llm-wiki/mcp-server/src/server.js"
      ],
      "env": {
        "LLM_WIKI_API_URL": "http://localhost:19828"
      }
    }
  }
}

Usage in OpenCode

The MCP server is automatically configured via opencode.jsonc in the project root:

{
  "mcp": {
    "servers": {
      "llm-wiki": {
        "command": "node",
        "args": ["node_modules/@opencode-llm-wiki/mcp-server/src/server.js"],
        "env": {
          "LLM_WIKI_API_URL": "http://localhost:19828"
        }
      }
    }
  }
}

Start the Backend Server

Before using the MCP server, make sure the OpenCode LLM Wiki backend is running:

# Clone the main repository
git clone https://github.com/Moowningstar/opencode-llm-wiki.git
cd opencode-llm-wiki

# Build and run the Rust backend
cargo run --bin llm-wiki-server -- --port 19828

The MCP server communicates with the backend via HTTP API.

🛠️ Available MCP Tools

The server exposes 11 tools for interacting with your knowledge base:

Core Tools

| Tool | Description | |------|-------------| | wiki_read | Read a wiki page by path | | wiki_list | List all wiki pages | | wiki_search | Search pages by keyword | | wiki_query_with_context | Intelligent context injection (keyword + vector search) | | wiki_ingest | Ingest documents into the knowledge graph |

Graph Tools

| Tool | Description | |------|-------------| | wiki_get_graph | Get knowledge graph data (nodes and edges) | | wiki_graph_insights | Analyze graph structure (isolated pages, bridges, statistics) | | wiki_deep_research | Multi-hop reasoning with graph traversal and semantic search |

Metadata Tools

| Tool | Description | |------|-------------| | wiki_get_index | Get index.md (content catalog) | | wiki_get_overview | Get overview.md (global summary with graph statistics) | | wiki_get_purpose | Get purpose.md (wiki goals and scope) |

📖 Tool Examples

Read a Page

{
  "tool": "wiki_read",
  "arguments": {
    "path": "entities/gpt-4.md"
  }
}

Search with Context

{
  "tool": "wiki_query_with_context",
  "arguments": {
    "query": "transformer architecture",
    "max_tokens": 4000
  }
}

Deep Research

{
  "tool": "wiki_deep_research",
  "arguments": {
    "query": "attention mechanisms in neural networks",
    "max_depth": 3,
    "max_results": 10
  }
}

Graph Insights

{
  "tool": "wiki_graph_insights",
  "arguments": {
    "analysis_type": "all"  // Options: "isolated", "bridges", "stats", "all"
  }
}

Ingest Documents

{
  "tool": "wiki_ingest",
  "arguments": {
    "path": "docs/architecture.md",
    "recursive": false
  }
}

🏗️ Architecture

The MCP server acts as a bridge between AI agents (like Claude Desktop) and the OpenCode LLM Wiki backend:

┌─────────────────┐
│  Claude Desktop │
│   (AI Agent)    │
└────────┬────────┘
         │ MCP Protocol (stdio)
         ↓
┌─────────────────┐
│   MCP Server    │
│  (Node.js)      │
└────────┬────────┘
         │ HTTP API
         ↓
┌─────────────────┐
│  Rust Backend   │
│  (Port 19828)   │
│                 │
│  • RuVector DB  │
│  • Graph Algos  │
│  • Vector Search│
└─────────────────┘

Components

• MCP Server (src/server.js): Implements Model Context Protocol, exposes 11 tools
• Core API Client (src/lib/core-api-client.js): HTTP client for backend communication
• Backend Server: Rust service providing vector search, graph algorithms, and storage

🔧 Development

Run Server Manually

cd src-mcp
LLM_WIKI_API_URL=http://localhost:19828 node src/server.js

The server runs in stdio mode for MCP protocol communication.

Debug Mode

Set the DEBUG environment variable to see detailed logs:

DEBUG=llm-wiki:* LLM_WIKI_API_URL=http://localhost:19828 node src/server.js

Testing Tools

You can test individual tools using the MCP inspector or by sending JSON-RPC requests:

# Example: Test wiki_list tool
echo '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"wiki_list","arguments":{}}}' | node src/server.js

📝 Configuration

Environment Variables

• LLM_WIKI_PROJECT - Project root path (default: ~/llm-wiki-projects/default)
• LLM_WIKI_API_URL - API server URL (optional, for future HTTP mode)

Project Structure

my-wiki/
├── .raw/              # Drop files here for auto-ingest
├── .wiki/
│   ├── entities/      # Named things (models, companies, people)
│   ├── concepts/      # Ideas, techniques, phenomena
│   ├── sources/       # Source summaries
│   ├── queries/       # Research questions
│   ├── index.md       # Content catalog (auto-maintained)
│   └── overview.md    # Global summary (auto-generated)
├── purpose.md         # Wiki goals and scope
├── schema.md          # Structure rules
└── .llm-wiki/         # Internal data (database, cache)

🔗 Integration with OpenCode

How It Works

1. OpenCode starts → Reads opencode.jsonc
2. MCP server launches → node src-mcp/src/server.js
3. Server initializes → Connects to wiki project at ${workspaceFolder}
4. Tools available → OpenCode can call any of the 10 MCP tools
5. File watching → Server monitors .raw/ and .wiki/ for changes

Usage in OpenCode

When you ask OpenCode a question about your knowledge base:

You: "What do I know about transformer architecture?"

OpenCode internally calls:
1. wiki_query_with_context(query="transformer architecture")
2. Receives relevant wiki pages with context
3. Answers your question with citations

🚨 Troubleshooting

Server Won't Start

# Check Node.js version (requires >= 18)
node --version

# Reinstall dependencies
cd src-mcp
rm -rf node_modules package-lock.json
npm install

No Wiki Pages Found

# Verify project structure
ls .wiki/

# Check if pages exist
ls .wiki/entities/
ls .wiki/concepts/

Vector Search Not Working

Vector search requires embeddings. If you see warnings about embedding generation:

1. Check if OpenAI API key is configured (for embedding generation)
2. Or use keyword-only search (works without embeddings)

📚 Related Documentation

• MCP Server Guide - Detailed architecture and usage
• Project README - Main project documentation
• Model Context Protocol - MCP specification

📄 License

GPL-3.0