Asset-Aware MCP

🏗️ Asset-Aware ETL for AI Agents - Precise PDF decomposition into structured assets (Tables, Figures, Sections)

🆕 What's New in v0.2.10

• Modular Architecture: Server refactored from 2122 → 31 lines (thin entry point)
• 34 MCP Tools across 5 modules (Document, Section, Job, Knowledge, Table)
• 12 MCP Resources across 2 modules (Document, Table)
• Bug Fixes: use_marker async mode, list_documents filtering, image overlap detection

🌟 Core Concept: Asset-Aware ETL

This extension provides a sophisticated ETL (Extract, Transform, Load) Pipeline for AI Agents. Instead of feeding raw text to an LLM, it decomposes documents into a structured "Map" (Manifest), allowing Agents to precisely retrieve what they need.

The Workflow:

1. 📥 Ingest (ETL): Agent provides a local PDF path.
2. ⚙️ Process: MCP Server reads the file using PyMuPDF, separating Text, Tables, and Figures (with page numbers).
3. 🗺️ Manifest: Generates a structured JSON "Map" of all assets.
4. 📤 Fetch: Agent "looks at the map" and fetches specific objects (e.g., "Table 1" or "Figure 2") as clean Markdown or Base64 images.

✨ Features

• 📄 Dual-Engine PDF ETL:
  • PyMuPDF (default) - Fast extraction (~50MB dependency)
  • Marker (optional, use_marker=True) - High-precision with blocks.json containing bbox coordinates
• 🧭 Section Navigation: Dynamic hierarchy section tree with 4 tools for browsing, searching, and block extraction
• 🔄 Async Jobs: Track progress for large document batches with Job IDs.
• 🗺️ Document Manifest: A structured index that lets Agents "see" document structure before reading.
• 🖼️ Visual Assets: Extract figures as Base64 images for Vision-capable Agents.
• 📊 A2T (Anything to Table): 19 tools for creating, editing, and exporting professional Excel tables
• 🧠 Knowledge Graph: Cross-document insights powered by LightRAG.
• 🔌 MCP Native: Seamless integration with VS Code Copilot Chat and Claude.
• 🏠 Local-First: Optimized for Ollama (local LLM) but supports OpenAI.

🚀 Quick Start

1. Install Prerequisites

# Install Ollama (for local LLM)
curl -fsSL https://ollama.com/install.sh | sh

# Pull required models
ollama pull qwen2.5:7b
ollama pull nomic-embed-text

2. Install Extension

1. Open VS Code
2. Go to Extensions (Ctrl+Shift+X)
3. Search for "Asset-Aware MCP"
4. Click Install

3. Run Setup Wizard

1. Open Command Palette (Ctrl+Shift+P)
2. Run Asset-Aware MCP: Setup Wizard
3. Follow the prompts to configure your .env file.

📖 Usage (Agent Flow)

1. Ingest a Document (ETL)

In Copilot Chat, tell the agent to process a file: @workspace Use ingest_documents to process ./papers/study_01.pdf

2. Check Progress

For large files, check the job status: @workspace get_job_status("job_id_here")

3. Inspect the Map

The agent will first look at the manifest to see what's inside: @workspace What tables are available in doc_study_01?

4. Fetch Specific Assets

The agent retrieves exactly what it needs: @workspace Fetch Table 1 from doc_study_01 @workspace Show me Figure 2.1 (the study flow diagram)

⚙️ Configuration

🔧 Commands

🛠️ Troubleshooting & Debugging

If the extension fails to start or the MCP server doesn't appear:

1. Check VS Code Version: Ensure you are using VS Code 1.96.0 or newer.
2. Check Dependencies: Run Asset-Aware MCP: Check System Dependencies from the command palette.
3. Inspect Logs:
   • Open Output panel (Ctrl+Shift+U).
   • Select Asset-Aware MCP from the dropdown to see extension logs.
   • Select Asset-Aware MCP Dependencies to see dependency check results.
4. Development Mode:
   • Clone the repo.
   • Open vscode-extension folder.
   • Run npm install.
   • Press F5 to launch the Extension Development Host.

📚 MCP Tools (34 total)

Document ETL (5)

Section Navigation (4)

Job Management (4)

Knowledge Graph (2)

A2T - Anything to Table (19)

🔗 Links

• GitHub Repository
• PyPI Package
• Technical Specification

📝 License

Apache-2.0