What is MCP, and why did DiligenceVault build it?

Naman:To kick things off simply: DiligenceVault has spent over a decade perfecting an enterprise SaaS platform. Why build an MCP server now, and what does it actually do for someone who isn't deep in AI terminology?

Rimsha:Think of our standard web interface as a digital workspace built for human eyes and human fingers - buttons, dropdowns, and dashboards. The Model Context Protocol (MCP) is an open standard that creates an equivalent workspace for AI assistants like Claude. Instead of forcing an analyst to manually log in, navigate five menus, download two spreadsheets, and paste them into an AI chat, MCP allows the AI model to securely query DiligenceVault directly behind the scenes.

Shrey:From an engineering standpoint, the protocol itself is just a standardized translation bridge. The opportunity was that our web app has ten years of built-in business rules, role permissions, and edge cases. When you open a door for an AI agent, you can't just point it at a raw database. You have to teach it how to interact with your platform safely, cleanly, and predictably and our accrued knowledge makes our MCP architecture more powerful.

What does MCP change for an investment analyst?

Naman:How does all of this translate to the business user - the investment analyst or research director working against a deadline?

Rimsha:It solves the "five-tab problem." Today, answering a question like "How did Manager X's response on cyber risk change between their 2024 and 2026 questionnaires, and did our risk team flag any follow-ups?" requires an analyst to open multiple tabs, pull up historical archives, download audit notes, and manually compare them.

With DiligenceVault MCP, the analyst asks that exact question in plain English inside their preferred AI workspace. The model uses the MCP server to pull both sets of historical records, verify the internal notes, and compare the responses in seconds, without the analyst exporting a single file.

How does MCP extend DV Assist beyond the platform?

Naman:How does MCP fit alongside DV Assist? When would a client use each?

Rimsha:DV Assist is our native assistant built directly into the DiligenceVault application, designed for users working inside our platform who want to query documents, respond to questionnaires, review questionnaires, generate memos with robust governance and reusable prompts right where their data lives. Model Context Protocol (MCP) extends that intelligence outward, allowing clients to bring DiligenceVault's secure first party data and context directly into the external AI tools and enterprise workspaces their teams already use, such as Claude. A client uses DV Assist when they are operating within our core platform, and they use MCP when they want to query DiligenceVault seamlessly from their preferred external enterprise AI environment.

How does MCP respect existing data permissions?

Naman:When Chief Compliance Officers (CCOs) and risk committees hear "AI connecting directly to our financial data," their alarms go off. Your first design rule was: "Your MCP server should never be a back door." How do you reassure a CCO that their data remains locked down?

Shrey:We established strict permission parity from day one. Our MCP server does not have its own "super-admin" access key, nor does it maintain a secondary permissions database that could fall out of sync. When an AI client makes a request through MCP on behalf of an analyst, it passes that exact analyst's authentication credentials straight to our core backend. If a public markets analyst cannot see a manager's private fund on their web screen, the AI cannot see it through MCP either. Period.

Rimsha:That gives compliance leaders immense peace of mind. You don't need to configure a new set of data governance rules for AI. Your existing information barriers, NDA firewalls, role-based access controls (RBAC), and multi-tenant security automatically apply.

Can AI change data without human approval?

Naman:What about operational risk? Can an AI agent hallucinate an instruction and accidentally overwrite a live due diligence questionnaire (DDQ) or trigger an unapproved email to a fund manager?

Rimsha:Absolutely not, because of how we separate reading data from modifying data. We adopted a strict read-first rollout. The AI is empowered to search, cross-reference, summarize, and retrieve diligence intelligence. For any action that changes state - like submitting an evaluation or sending an RFP - we enforce a mandatory "human-in-the-loop" verification step. The AI can draft the action, but a human must review and click "Approve" in the core platform before anything goes live.

Does MCP require rebuilding existing systems?

Naman:Let's look at this from a CTO's perspective. Enterprise tech leaders are tired of hype cycles that demand complete architecture rewrites. Did your team have to rebuild the backend to make this work?

Shrey:Not at all. That was another of our design rules: most of your APIs can stay as they are. We didn't rewrite our core business logic or replace our services. Instead, the MCP server acts as an intelligent adapter.

A traditional web API often returns large payloads with styling data, pagination links, and internal status flags that a human browser needs. If you dump all of that raw data into an LLM, you burn through context tokens, increase latency, and run up inference costs. Our MCP layer trims the fat: it takes the output from our existing APIs, strips out UI-specific metadata, reshapes the information into clean, token-efficient JSON, and sends the LLM only the precise fields it needs to answer the question.

Why do fewer tools make AI more accurate?

Naman:Engineering teams building their first MCP server often fall into the trap of turning every existing API endpoint into an AI tool. You advised against that. Why is a large tool inventory dangerous?

Shrey:It directly degrades accuracy. If you give an AI model excessive hyper-specific tools with overlapping purposes, you dramatically increase the likelihood of hallucination. The model gets confused about which tool to select, passes parameters to the wrong endpoint, or gets caught in an execution loop.

Rimsha:We focused on composite tools designed around user intent. Instead of giving the AI three separate tools - GetManagerProfile, GetHistoricalTrackRecord, and GetComplianceFlags - we combined them into a single, cohesive diligence tool. This keeps the model's decision-making surface small and reliable, which is critical when clients are making high-stakes allocation decisions.

What did the launch teach about rate limits and tool descriptions?

Naman:Shrey, what technical lessons should other engineering leaders take away from our launch, especially around rate limits and where human review is still needed?

Shrey:Two big operational realities:

First, AI calls APIs very differently than humans do. A human clicks a button and waits. An LLM reasons in iterative loops - it might run a search, inspect the result, realize it needs more context, and trigger three more automated calls in five seconds. If your API gateway has rigid, legacy rate limits, the AI workflow breaks instantly. You need flexible burst limits and real-time observability on tool execution.

Second, don't trust an AI model to write its own tool documentation. The descriptions you write for your tools and their parameters are effectively the "code" the LLM reads to decide how to act. A single ambiguous property description can cause a model to misinterpret an entire workflow. Human engineers who understand ten years of institutional edge cases must scrutinize every description line by line.

How should leadership teams approach MCP?

Naman:What is the final word for leadership teams evaluating whether to bring MCP into their enterprise product roadmap?

Rimsha:Launch day is an educated hypothesis, not the finish line. Don't wait for the ecosystem to achieve perfection before you begin. Define strict security boundaries, start with a handful of high-value read tools, watch how real users query your system, and let client feedback guide your iterations.