Production-ready corpus analysis for the age of AI
AI agents running advanced grammatical queries via the Model Context Protocol
Context-Fabric brings corpus analysis into the AI era. Built on the proven Text-Fabric data model, it introduces a memory-mapped architecture enabling parallel processing for production deployments—REST APIs, multi-worker services, and AI agent tools via MCP.
- Built for Production — Memory-mapped arrays enable true parallelization. Multiple workers share data instead of duplicating it.
- AI-Native — MCP server exposes corpus operations to Claude, GPT, and other LLM-powered tools.
- Powerful Data Model — Standoff annotation, graph traversal, pattern search, and arbitrary feature annotations.
- Dramatic Efficiency — 3.5x faster loads, 65% less memory in single process, 62% less with parallel workers.
Context-Fabric includes cfabric-mcp, a Model Context Protocol server that exposes corpus operations to AI agents:
# Start the MCP server
cfabric-mcp --corpus /path/to/bhsa
# Or with SSE transport for remote clients
cfabric-mcp --corpus /path/to/bhsa --sse 8000The server provides 10 tools for discovery, search, and data access—designed for iterative, token-efficient agent workflows.
Text-Fabric loads entire corpora into memory—effective for single-user research, but each parallel worker duplicates that memory footprint. Context-Fabric's memory-mapped arrays change the equation:
| Scenario | Memory Reduction |
|---|---|
| Single process | 65% less |
| 4 workers (spawn) | 62% less |
| 4 workers (fork) | 62% less |
Mean reduction across 10 corpora. Memory measured as total RSS after loading from cache.
# Core library
pip install context-fabric
# With MCP server
pip install context-fabric[mcp]from cfabric.core import Fabric
# Load a corpus
CF = Fabric(locations='path/to/corpus')
api = CF.load('feature1 feature2')
# Navigate nodes
for node in api.N.walk():
print(api.F.feature1.v(node))
# Traverse structure
embedders = api.L.u(node) # nodes containing this node
embedded = api.L.d(node) # nodes within this node
# Search patterns
results = api.S.search('''
clause
phrase function=Pred
word sp=verb
''')| API | Purpose |
|---|---|
| N | Walk nodes in canonical order |
| F | Access node features |
| E | Access edge features |
| L | Navigate locality (up/down the hierarchy) |
| T | Retrieve text representations |
| S | Search with structural templates |
Context-Fabric trades one-time compilation cost for dramatic runtime efficiency. Compile once, benefit forever.
| Metric | Mean Improvement |
|---|---|
| Load time | 3.5x faster |
| Memory (single) | 65% less |
| Memory (spawn) | 62% less |
| Memory (fork) | 62% less |
Mean across 10 corpora. The larger cache enables memory-mapped access—no deserialization, instant loads, shared memory across workers.
Run benchmarks yourself:
pip install context-fabric[benchmarks]
cfabric-bench memory --corpus path/to/corpus| Package | Description |
|---|---|
| context-fabric | Core graph engine |
| cfabric-mcp | MCP server for AI agents |
| cfabric-benchmarks | Performance benchmarking suite |
If you use Context-Fabric in your research, please cite:
Kingham, Cody. "Carrying Text-Fabric Forward: Context-Fabric and the Scalable Corpus Ecosystem." January 2026.
Context-Fabric by Cody Kingham, built on Text-Fabric by Dirk Roorda.
MIT