WSO2 Docs MCP Server

A production-ready Model Context Protocol (MCP) server that provides AI assistants (Claude Desktop, Claude Code, Cursor, VS Code) with semantic search over WSO2 documentation via Retrieval-Augmented Generation (RAG).

Under the hood, it uses a blazing-fast dual-ingestion engine:

• GitHub Native: Fetches raw Markdown directly from WSO2's public GitHub repositories via the Git Trees API (avoids web-scraping noise and rate limits)
• Web Crawl Fallback: For products without dedicated GitHub docs repos (like the WSO2 Library)

Architecture

flowchart TD
    %% Styling
    classDef github fill:#1f2328,color:#fff,stroke:#e1e4e8
    classDef default fill:#0969da,color:#fff,stroke:#0969da
    classDef db fill:#218bff,color:#fff,stroke:#218bff
    classDef web fill:#0969da,color:#fff,stroke:#0969da

    subgraph Sources ["Information Sources"]
        GH["GitHub Repos\n(wso2/docs-apim, etc.)"]:::github
        Web["WSO2 Websites\n(ballerina.io, lib)"]:::web
    end

    subgraph Ingestion ["Dual-Ingestion Pipeline"]
        A["GitHubDocFetcher\n(Git Trees API)"] 
        B["DocCrawler\n(HTML scraping)"]
        
        C["MarkdownParser\n(Front-matter & Heads)"]
        D["DocParser\n(Cheerio HTML parsing)"]
        
        E["DocChunker\n(Semantic Splitting)"]
    end

    subgraph Embedding ["Vectorization & Storage"]
        F["EmbedderFactory\n(Ollama / HuggingFace)"]
        G[("pgvector\n(PostgreSQL)")]:::db
    end

    %% Flow
    GH --> A
    Web --> B
    
    A -->|"Raw .md"| C
    B -->|"Clean HTML"| D
    
    C -->|"ParsedSection[]"| E
    D -->|"ParsedSection[]"| E
    
    E -->|"Tokens/Chunks"| F
    F -->|"768-dim Vectors"| G

Documentation Sources

| Product | URL | |---|---| | API Manager | https://apim.docs.wso2.com | | Micro Integrator | https://mi.docs.wso2.com/en/4.4.0 | | Choreo | https://wso2.com/choreo/docs | | Ballerina | https://ballerina.io/learn | | Ballerina Integrator | https://bi.docs.wso2.com | | WSO2 Library | https://wso2.com/library |

Prerequisites

• Node.js ≥ 20
• Docker (for pgvector)
• Embeddings — no API key required by default:
  • Ollama (recommended) — runs locally, model auto-downloaded on first run
  • If Ollama is not running, the server automatically falls back to HuggingFace ONNX (in-process, also downloads automatically)
  • Cloud providers are also supported: OpenAI, Google Gemini, Voyage AI

Quick Start

Choose the setup path that fits your use case:

• Install from npm — simplest, no cloning required
• Clone and build — for development or contributions

Install from npm

Install the package globally to get the wso2-docs-mcp-server, wso2-docs-crawl, and wso2-docs-migrate commands available system-wide:

npm install -g wso2-docs-mcp-server

Prefer no global install? You can use npx wso2-docs-mcp-server, npx wso2-docs-crawl, and npx wso2-docs-migrate in every step below — just replace the bare command with its npx equivalent.

1. Start pgvector

Download the docker-compose.yml and start the database:

curl -O https://raw.githubusercontent.com/iamvirul/wso2-docs-mcp-server/main/docker-compose.yml
docker compose up -d

2. Start Ollama (optional but recommended)

Install Ollama and pull the default embedding model:

ollama pull nomic-embed-text
ollama serve

No Ollama? Skip this step. The server automatically falls back to HuggingFace ONNX — model downloads on first use with no extra setup.

3. Run database migration

DATABASE_URL="postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs" \
  wso2-docs-migrate

Run migration again whenever you change EMBEDDING_DIMENSIONS (i.e. switch embedding provider). The script detects and handles dimension changes automatically.

4. Index WSO2 documentation

# Index all products (first run downloads the embedding model automatically)
DATABASE_URL="postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs" \
  wso2-docs-crawl

# Index a single product (faster, great for testing)
DATABASE_URL="postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs" \
  wso2-docs-crawl --product ballerina --limit 20

# Force re-index even unchanged pages
DATABASE_URL="postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs" \
  wso2-docs-crawl --force

Available product IDs: apim, mi, choreo, ballerina, bi, library

5. Configure your AI client

The MCP server is launched on demand by your AI client — no background process needed.

Claude Desktop — edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "wso2-docs": {
      "command": "wso2-docs-mcp-server",
      "env": {
        "DATABASE_URL": "postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs",
        "EMBEDDING_PROVIDER": "ollama"
      }
    }
  }
}

Claude Code — run once in your terminal:

claude mcp add wso2-docs \
  --transport stdio \
  -e DATABASE_URL="postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs" \
  -e EMBEDDING_PROVIDER="ollama" \
  -- wso2-docs-mcp-server

# Verify
claude mcp list

Cursor — create .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "wso2-docs": {
      "command": "wso2-docs-mcp-server",
      "env": {
        "DATABASE_URL": "postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs",
        "EMBEDDING_PROVIDER": "ollama"
      }
    }
  }
}

VS Code — create .vscode/mcp.json:

{
  "servers": {
    "wso2-docs": {
      "type": "stdio",
      "command": "wso2-docs-mcp-server",
      "env": {
        "DATABASE_URL": "postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs",
        "EMBEDDING_PROVIDER": "ollama"
      }
    }
  }
}

Using npx instead of global install? Replace "command": "wso2-docs-mcp-server" with "command": "npx" and add "args": ["-y", "wso2-docs-mcp-server"].

Cloud embedding provider? Add the key to env, e.g. "EMBEDDING_PROVIDER": "openai", "OPENAI_API_KEY": "sk-...".

Clone and build

1. Clone and install

git clone https://github.com/iamvirul/wso2-docs-mcp-server.git
cd wso2-docs-mcp-server
npm install

2. Start Ollama (optional but recommended)

Install Ollama and start it:

ollama serve

No Ollama? Skip this step. The server detects Ollama is not running and automatically falls back to HuggingFace ONNX inference — the model downloads on first use with no extra setup.

3. Configure environment

cp .env.example .env
# Defaults work out of the box with Ollama.
# Only edit if using a cloud provider (OpenAI / Gemini / Voyage).

4. Start pgvector

docker compose up -d
# pgAdmin available at http://localhost:5050 ([email protected] / admin)

5. Run database migration

npm run db:migrate

Note: Run migration again whenever you change EMBEDDING_DIMENSIONS (i.e. switch embedding provider). The script detects and handles dimension changes automatically.

6. Index documentation

# Index all products
# On first run the embedding model is downloaded automatically (Ollama or HuggingFace)
npm run crawl

# Index a single product (faster, great for testing)
npm run crawl -- --product ballerina --limit 20

# Force re-index even unchanged pages
npm run crawl -- --force

7. Build and start the MCP server

npm run build
npm start

For development (no build step):

npm run dev

8. Configure your AI client

Replace /ABSOLUTE/PATH/TO/wso2-docs-mcp-server with your actual clone path.

Claude Desktop — edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "wso2-docs": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/TO/wso2-docs-mcp-server/dist/src/index.js"],
      "env": {
        "DATABASE_URL": "postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs",
        "EMBEDDING_PROVIDER": "ollama"
      }
    }
  }
}

Claude Code:

claude mcp add wso2-docs \
  --transport stdio \
  -e DATABASE_URL="postgresql://wso2mcp:wso2mcp@localhost:5432/wso2docs" \
  -e EMBEDDING_PROVIDER="ollama" \
  -- node "/ABSOLUTE/PATH/TO/wso2-docs-mcp-server/dist/src/index.js"

# Verify
claude mcp list

See config-examples/claude_code.sh for a convenience script.

Cursor — create .cursor/mcp.json — see config-examples/cursor_mcp.json.

VS Code — create .vscode/mcp.json — see config-examples/vscode_mcp.json.

MCP Tools

| Tool | Description | |---|---| | search_wso2_docs | Semantic search across all products. Optional product and limit filters. | | get_wso2_guide | Search within a specific product (apim, mi, choreo, ballerina, bi, library). | | explain_wso2_concept | Broad concept search across all products, returns 8 top results. | | list_wso2_products | Returns all supported products with IDs and base URLs. |

Example response

[
  {
    "title": "Deploying WSO2 API Manager",
    "snippet": "WSO2 API Manager can be deployed in various topologies…",
    "source_url": "https://apim.docs.wso2.com/en/latest/install-and-setup/...",
    "product": "apim",
    "section": "Deployment Patterns",
    "score": 0.8712
  }
]

Local Embeddings

The default EMBEDDING_PROVIDER=ollama runs entirely on your machine with no API key. The startup sequence is:

Is Ollama running?
├── Yes → Is model present?
│         ├── Yes → Ready (instant)
│         └── No  → Pull via Ollama (streamed, runs once)
└── No  → Download ONNX model from HuggingFace Hub (~250 MB, cached after first run)
           and run inference in-process via @huggingface/transformers

Both paths use nomic-embed-text / Xenova/nomic-embed-text-v1 by default and produce identical 768-dim vectors, so you can switch between them without re-indexing.

Hardware acceleration (HuggingFace ONNX fallback)

When Ollama is not available, the server auto-detects the best compute backend:

| Machine | Detection | ONNX dtype | Batch size | Throughput | |---|---|---|---|---| | Apple Silicon (M1/M2/M3/M4) | process.arch === 'arm64' | q8 INT8 | 32 | ~9 ms/chunk | | NVIDIA GPU | nvidia-smi probe | fp32 | 64 | GPU-dependent | | All others | fallback | q8 INT8 | 16 | ~10 ms/chunk |

Why q8 on Apple Silicon instead of CoreML/Metal? CoreML compiles Metal shaders on first use (~20 min cold-start). For the typical chunk sizes produced by this server (6–20 chunks per page), the CPU↔GPU transfer overhead eliminates any inference gain. INT8 quantized inference on ARM NEON SIMD is consistently ~100× faster than fp32 CPU with zero cold-start cost.

Benchmark (Apple M-chip, Xenova/nomic-embed-text-v1):

fp32 CPU (before): ~1,000 ms/chunk   (68 chunks ≈ 68 s of embedding)
q8  ARM NEON:          ~9 ms/chunk   (68 chunks ≈  0.6 s of embedding)  ← ~100× speedup

Note: For small crawls (≤ 10 pages) total wall-clock time is dominated by network I/O (HTTPS fetches to docs sites), so the end-to-end improvement is modest. The embedding speedup becomes significant at scale — crawling 500+ pages where embedding previously accounted for hours of runtime. For best crawl performance, run Ollama (ollama serve) which parallelises inference natively and has no per-chunk overhead.

Environment Variables

Core

| Variable | Default | Description | |---|---|---| | DATABASE_URL | — | PostgreSQL connection string (required) | | EMBEDDING_PROVIDER | ollama | ollama | openai | gemini | voyage | | EMBEDDING_DIMENSIONS | 768 | Must match model output dimensions | | CRAWL_CONCURRENCY | 5 | Concurrent HTTP requests during crawl | | CHUNK_SIZE | 800 | Approximate tokens per chunk | | CHUNK_OVERLAP | 100 | Overlap tokens between chunks | | CACHE_TTL_SECONDS | 3600 | In-memory query cache TTL | | TOP_K_RESULTS | 10 | Default search result count |

Ollama (default)

| Variable | Default | Description | |---|---|---| | OLLAMA_BASE_URL | http://localhost:11434 | Ollama server URL | | OLLAMA_EMBEDDING_MODEL | nomic-embed-text | Model pulled and used via Ollama | | HUGGINGFACE_EMBEDDING_MODEL | Xenova/nomic-embed-text-v1 | ONNX fallback when Ollama is not running |

Cloud providers

| Variable | Default | Description | |---|---|---| | OPENAI_API_KEY | — | Required if EMBEDDING_PROVIDER=openai | | OPENAI_EMBEDDING_MODEL | text-embedding-3-small | OpenAI model | | GEMINI_API_KEY | — | Required if EMBEDDING_PROVIDER=gemini | | GEMINI_EMBEDDING_MODEL | text-embedding-004 | Gemini model | | VOYAGE_API_KEY | — | Required if EMBEDDING_PROVIDER=voyage | | VOYAGE_EMBEDDING_MODEL | voyage-3 | Voyage model |

Embedding dimension reference

| Provider | Model | Dimensions | |---|---|---| | Ollama / HuggingFace | nomic-embed-text / Xenova/nomic-embed-text-v1 | 768 (default) | | Ollama / HuggingFace | mxbai-embed-large / Xenova/mxbai-embed-large-v1 | 1024 | | Ollama / HuggingFace | all-minilm / Xenova/all-MiniLM-L6-v2 | 384 | | OpenAI | text-embedding-3-small | 1536 | | OpenAI | text-embedding-3-large | 3072 | | Gemini | text-embedding-004 | 768 | | Voyage | voyage-3 | 1024 | | Voyage | voyage-3-lite | 512 |

Scheduled Re-indexing

# Run a one-off re-index (checks hashes, skips unchanged pages)
npm run reindex

# Or from the project directory using node-cron (runs daily at 2 AM)
DATABASE_URL=... node -e "
  const { ReindexJob } = require('./dist/jobs/reindexDocs');
  const job = new ReindexJob();
  job.initialize().then(() => job.scheduleDaily());
"

Project Structure

src/
  config/          env.ts · constants.ts
  vectorstore/     pgvector.ts · schema.sql
  ingestion/       crawler.ts · parser.ts · githubFetcher.ts · markdownParser.ts · chunker.ts · embedder.ts
  server/          mcpServer.ts · toolRegistry.ts
  jobs/            reindexDocs.ts
  index.ts
scripts/
  crawl.ts         CLI ingestion pipeline
  migrate.ts       Dynamic schema migration
config-examples/   claude_desktop.json · claude_code.sh · cursor_mcp.json · vscode_mcp.json
docker-compose.yml
.env.example

Development

# Type-check
npx tsc --noEmit

# Run crawl with tsx (no build needed)
npm run crawl -- --product ballerina --limit 5

# Run server in dev mode
npm run dev