Skip to main content

What is MCP?

The Model Context Protocol (MCP) is an open standard for connecting AI agents to external data sources and tools. Instead of pasting data into prompts, MCP lets agents call structured tools to query, retrieve, and act on data in real time. Claude, ChatGPT, Microsoft Copilot, and any MCP-compatible client can connect to an MCP server and use its tools as part of their reasoning.

Why MCP for drawings?

LLMs can reason about construction drawings — cross-referencing schedules, checking code compliance, coordinating across disciplines — but they need structured access to the parsed data. Raw PDFs are too large and too dense for a single prompt. A 200-page drawing set contains thousands of cross-references that span sheets and disciplines. The Drawing Index MCP gives agents composable query tools to traverse the knowledge graph interactively. The agent decides what to look up, when to drill deeper, and how to combine results. Bedrock provides the data; the agent provides the reasoning.

Design Principles

Composable Primitives

Each tool does one thing. The agent chains them to accomplish complex workflows like code compliance checks or RFI drafting. The MCP never imposes a workflow the agent can’t deviate from.

Context-Rich Responses

Every entity includes parent context. A feature comes with its block type and sheet number. A block comes with its sheet and drawing. No extra calls just to get human-readable context.

Progressive Disclosure

List operations return lightweight summaries by default. Heavy data (OCR text, schedule metadata, overlay changes) is gated behind include parameters. The agent pays the token cost only when it needs detail.

Explicit Async Boundary

Vision processing is async. The Compare tool returns a job_id. A separate PollJob tool polls status. Tools never block — the agent decides when and how often to poll.

Pass-Through Filters

Filters reference parent entities directly. The Query tool accepts pass-through filters like sheet_number, block_type, and block_identifier that join through parent entities. This avoids the “resolve then query” round-trip pattern.

Flat Tool Surface

Seven tools total. No nested namespaces, no tool discovery step. Every MCP-compatible agent sees all tools immediately on connection.

Tools

Seven tools organized in four categories:
ToolCategoryDescription
QueryDataGet, list, or filter any entity
SearchDataHybrid semantic + keyword search
AskDataNatural language Q&A with sources
CompareVisionAsync drawing comparison
ParseVisionAsync feature extraction
PollJobAsyncPoll job status and results
ViewImageContentStorage URI to signed URL
Across 9 reference workflows, agents average 2.2 tool calls per task. The pass-through filters and context-rich responses keep round-trips low.

Next Steps

Setup

Connect the Bedrock MCP server to Claude, ChatGPT, Microsoft Copilot, or your custom agent.

Tools Reference

Full parameter and response documentation for all seven tools.

Workflow Examples

Step-by-step tool call sequences for real construction queries.

Best Practices

Prompt tips, context management, and error handling.