Context-Fabric MCP Server

Query Context-Fabric corpora via MCP (Model Context Protocol).

Installation

pip install cfabric-mcp

Transports

The server supports three transport modes:

Usage

# stdio transport (Claude Desktop)
cfabric-mcp --corpus /path/to/bhsa

# SSE transport (Cursor, remote clients)
cfabric-mcp --corpus /path/to/bhsa --sse 8000

# Streamable HTTP transport (production)
cfabric-mcp --corpus /path/to/bhsa --http 8000

Options

--corpus PATH              # Corpus to load (required, repeatable)
--corpus name=PATH         # Named corpus
--sse PORT                 # SSE transport on port (endpoint: /sse)
--http PORT                # Streamable HTTP transport on port (endpoint: /mcp)
--host HOST                # Host to bind (default: 0.0.0.0)
--features "sp lex"        # Load only specific features
--verbose                  # Debug logging

Tools

The server exposes 11 MCP tools organized in layers for efficient discovery.

Note: A usage guide is automatically provided to clients via the MCP instructions field at connection time.

Discovery Tools

Search Tools

Search totals count all matches, including matches beyond the cursor cache's 10,000-result retention limit. Counts remain exact on cache hits. Statistics use all matches; when only a prefix is cached, statistics may rerun the query. Result and passage responses include cached_count and results_truncated to distinguish retained results from total_count. Pagination's has_more describes remaining cached results, so it becomes false at the cache boundary even when results_truncated is true. CSV exports honor their requested row limit independently of cache retention, rerunning a truncated cached query when more rows are needed.

Data Access Tools

Discovery Flow

The tools are designed for hierarchical, iterative discovery:

describe_corpus()
│
├─► list_features()                  → Browse all features
│       └─► describe_feature("sp")   → Full details + samples
│
├─► list_features(node_types=["word"])  → Filter by node type
│       └─► describe_feature(...)
│
└─► get_text_formats()               → When encoding matters

        ↓
    search()  → Execute queries

Typical workflow:

1. describe_corpus() - Get structure overview
2. list_features() or list_features(node_types=["word"]) - Browse available features
3. describe_feature("sp") - Deep dive into specific feature with sample values
4. search_syntax_guide() - Get search syntax help (section-based)
5. search() - Execute search queries

Search Return Types

Client Configuration

Claude Desktop (stdio)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "context-fabric": {
      "command": "/path/to/venv/bin/cfabric-mcp",
      "args": ["--corpus", "/path/to/bhsa"]
    }
  }
}

Cursor (SSE)

Start server:

cfabric-mcp --corpus /path/to/bhsa --sse 8000

Configure Cursor to connect to: http://localhost:8000/sse

Custom MCP Clients

For SSE transport, connect to http://host:port/sse For Streamable HTTP, connect to http://host:port/mcp

License

MIT