Ask an AI assistant to “add a user to all my ACC projects.”

With ChatGPT or Copilot, it will generate boilerplate code. You will spend hours debugging authentication flows, pagination logic, and undocumented API quirks. The assistant will hallucinate endpoints. You will fix them. It will forget your project IDs. You will paste them again.

With Claude + RAPS MCP tools, it actually does it. Right now. One sentence in, users added, confirmation returned. No code generated. No code debugged. The AI calls the tools directly because RAPS exposes 114 tools that AI assistants can invoke as native operations.

This post is about what those 114 tools are, what they enable, and why no one else in CAD/PLM has anything remotely close.

What MCP Is

Model Context Protocol is an open standard for connecting AI assistants to external tools. Think of it as function calling, but standardized — any MCP-compatible assistant (Claude Desktop, Cursor, Claude Code, and others) can discover and use MCP tools without custom integration code.

The protocol works in three steps:

An MCP server starts and declares its available tools
An AI assistant connects and reads the tool catalog
When a conversation requires live data or real actions, the assistant calls the appropriate tool

RAPS is the only production MCP server for any CAD/PLM platform. Full documentation is at /docs/mcp-server. The MCP specification itself lives at modelcontextprotocol.io.

The 114 Tools

Every tool maps to a real APS API operation. They are organized into 16 domains:

Domain	Tools	Examples
Authentication	4	login, logout, status, test
Object Storage (OSS)	8	upload, download, list, delete, copy, signed URL, info, URN
Buckets	4	create, list, get, delete
Data Management	10	hub list/info, project list/info, folder contents/create, item info/versions/create/rename
Model Derivative	3	translate start, translate status, formats
Design Automation	5	engines, appbundles, activities, workitem create, workitem status
ACC Issues	6	list, get, create, update, comment add, comment delete
ACC RFIs	4	list, get, create, update
ACC Assets	4	list, get, create, update (delete)
ACC Submittals	3	list, create, update
ACC Checklists	3	list, create (from template), update
Account Admin	8	user add/remove/update role, project list/create/update/archive, folder permissions
Webhooks	6	create, list, get, update, delete, events
Reality Capture	6	create, process, status, result, delete, formats
Workflows	6	batch translate, compare versions, setup project, prepare for viewing, analyze model
Reports and Utilities	varies	issues summary, RFI summary, API request, skill info, pipeline ops

That is 114 tools, each with typed parameters, validation, rate-limit awareness, and error handling inherited from the same raps-kernel that powers the CLI.

What This Enables

The best way to understand the impact is to see real conversations.

Example 1 — Project Setup

You: Set up a new project called “Hospital Wing C”, add the structural team, and create the standard folder structure.

The AI calls project_create, then project_user_add five times (one per team member), then folder_create eight times for the standard directory tree. Total wall time: under 30 seconds. Manual equivalent: 15-20 minutes of clicking through the ACC admin UI.

Example 2 — Issue Triage

You: Show me all critical issues in the Hospital project that are overdue.

The AI calls issue_list with status and priority filters, formats the results into a table with assignees and due dates. No Postman. No GraphQL playground. No token management.

Example 3 — Translation Pipeline

You: Upload this Revit file, translate it to SVF2, and tell me when it’s ready.

The AI calls object_upload to push the file to OSS, translate_start to kick off the Model Derivative job, then polls translate_status until the job completes. You get a progress update in plain language.

Example 4 — Bulk Admin

You: Remove Sarah from all projects — she is moving to a different division.

The AI calls admin_user_remove with the account ID and Sarah’s email. One sentence, one operation, applied across every project in the account. The alternative is navigating to each project individually in the web UI.

Example 5 — Quality Check

You: List all open checklists for the Hospital project and summarize which ones are behind schedule.

The AI calls acc_checklists_list, filters for open items, and presents a summary with completion percentages and overdue flags.

The Competitive Landscape

Here is every known MCP integration for CAD/PLM platforms as of March 2026:

Vendor	MCP Server	Tools	Status
Autodesk (official)	None	0	No public plans
Petr Broz (community)	aps-mcp-server	~15	Basic, read-only, experimental
RAPS	raps mcp	114	Production, hosted at mcp.rapscli.xyz, paid tiers
PTC Onshape	None	0	—
Dassault 3DEXPERIENCE	None	0	—
Siemens Teamcenter	None	0	—
Trimble	None	0	—

RAPS is the only production MCP integration for any CAD or PLM platform. By an order of magnitude in tool count against the nearest community alternative.

Architecture

Starting the MCP server takes one command:

# Local mode — connects to your local raps installation
raps mcp serve

# Or use the hosted version
# Configure mcp.rapscli.xyz in Claude Desktop, Cursor, or any MCP client

Claude Desktop configuration:

&#123;
  "mcpServers": &#123;
    "raps": &#123;
      "command": "raps",
      "args": ["mcp"],
      "env": &#123;
        "APS_CLIENT_ID": "your_client_id",
        "APS_CLIENT_SECRET": "your_client_secret"
      &#125;
    &#125;
  &#125;
&#125;

Under the hood:

Built on rmcp 1.1 — the Rust MCP library, compiled into the same binary as the CLI
Same raps-kernel core as the CLI — identical auth, rate limiting, retry logic, and error handling
20 concurrent operations max — tuned to respect APS rate limits without tripping 429s
Sensitive data scrubbed from all responses — no tokens, secrets, or credentials leak to the AI
Clients cached for performance — created on-demand for specialized APIs, reused across tool calls

The MCP server is not a wrapper around the CLI. It shares the same Rust kernel. Every improvement to the CLI automatically improves the MCP tools, and vice versa.

Why This Matters for AEC

AEC professionals are not developers. They should not need to be.

The traditional path to APS automation requires learning OAuth 2.0, understanding REST API conventions, managing pagination, handling rate limits, and debugging opaque error codes. That path is a filter — it excludes the vast majority of people who would benefit from automation.

MCP tools change the economics. A project manager who can describe what they need in plain language now has access to the same automation that previously required a developer. The AI handles the API complexity. The professional handles the engineering decisions.

This is not about replacing developers. Developers will still build production pipelines, CI/CD workflows, and custom integrations. But the 80% of APS interactions that are ad hoc — checking status, pulling reports, onboarding users, triaging issues — those can now happen through conversation.

The convergence point is clear: AI assistants + domain-specific tools + production reliability = accessible automation. RAPS is where those three meet for AEC.

DevCon 2026 Session

This topic is the centerpiece of our DevCon Virtual session:

AI Pair-Assistant for APS Operators — 30-minute deep dive demonstrating real workflows with Claude + RAPS MCP tools. Live operations against a production ACC account. No slides, no mock data.

Read the full session preview at /blog/devcon-2026-ai-pair-assistant.

Try It

# Install RAPS
curl -fsSL https://rapscli.xyz/install.sh | bash

# Start the MCP server
raps mcp serve

Documentation: /docs/mcp-server

Next post in the pipeline: “From CLI to SaaS: How a Rust Monorepo Scales to Cloud” — the architecture that makes 114 tools possible from a single binary.