You give AI agents safe access to CDE data by exposing a governed, read-only copy of the project record through a small set of MCP tools that enforce permissions and log every call, instead of handing agents credentials to live CDE platforms. Bringing content into your own Azure tenancy first, through the Utopiaverse™, gives those tools one consistent, secure store to work from.
What AI agents are and why MCP matters
An AI agent is an AI application that can plan steps and call tools to complete a task, such as "prepare a summary of open design issues on the tunnel package". To do that, it needs tools that search documents, read registers and return results. The Model Context Protocol (MCP) is an open standard that defines how those tools are described and called, so one agent can use tools from many sources in a consistent way. Our article What Is MCP and How Does It Work With CDE Data? explains the basics.
The risks to manage
Agents are useful because they act on their own. That is also the risk. An agent connected directly to live CDEs with broad credentials could read content the user should not see, send too many requests to a vendor platform, or take an action that cannot be undone, such as changing a document status. It may also be influenced by instructions hidden in a document it reads, a risk known as prompt injection. A safe design limits what the agent can reach and what it can do.
A safe architecture
Replicate content from each CDE into your own Azure tenancy, so agents never touch vendor platforms directly.
Index the content with OCR and consistent metadata, so search results are reliable.
Write an MCP server that offers a small number of read-only tools over that index.
Pass the user's identity through to every tool call and filter results by their project permissions.
Log every request and response for audit and review.
Restrict which AI models and agent applications are approved to connect.
Designing the tools
Good MCP tools are narrow and predictable. Examples include:
search_documents: search by meaning across permitted projects, returning passages with document number, revision, platform and link.
get_current_revision: return the latest approved revision of a document by number.
list_transmittals: list transmittals by project, party and date range.
get_register_summary: return counts of documents by status for a project.
Each tool should return references to the source so the agent can cite them, and each should cap the volume of data returned. Avoid tools that write back to live platforms until the read-only tools are proven and the governance is agreed.
Guarding against prompt injection
Project documents come from many parties. Treat anything a document says as information only, never as a command. Keep tool outputs clearly separated from the agent's instructions, avoid giving agents tools that send messages or change records, and review logs for unusual patterns. Read-only access to a governed copy limits the harm an injected instruction can do.
Practical steps to get started
Pick one use case with clear value, such as answering design coordination questions on a single programme.
Agree permissions, logging and approved models with your security team.
Bring the relevant CDEs into your tenancy with CDE Sync™ or CDE Backup™.
Build and test two or three read-only MCP tools.
Pilot with a small group, review answers and citations, then expand.
Residency requirements can be checked against the global data region map, and Utopia Digital's security evidence, including SOC 2 Type II and Cyber Essentials, is in the Security Trust Portal.
AI agents and the Utopiaverse™
The Utopiaverse™ gives agents something safe and complete to work with. Content from Autodesk Forma, Oracle Aconex, Bentley ProjectWise, Procore, SharePoint and other CDEs sits in one connected record in your own Azure, with consistent metadata across every project. Your team can expose that record to approved agents through its own MCP server, while AI Insights gives everyone cited answers across the same record out of the box. The file viewer opens any document an agent or person cites, and dashboards show document control status across the portfolio.
The organisation gains visibility of every project and platform, with AI that works inside its own boundary and under its own rules. Our article on running AI across multiple CDEs in your own Azure covers the security model in more depth.
See the Utopiaverse™ in the monthly live demo, or talk to the team about AI agents on your project data.