Claude is, by most measures, the most capable reasoning model available for professional knowledge work. It can analyse financial statements, draft memos, compare deal terms, and synthesise research with a level of nuance that was not possible even twelve months ago. But out of the box, Claude has a fundamental limitation: it cannot see your data. It does not know what is in your CRM, what deals are in your pipeline, what your portfolio companies reported last quarter, or what your research team wrote about a sector three months ago.

This is the isolation problem. Claude is powerful, but it is operating in a vacuum. Ask it to draft an IC memo and it will produce something structurally sound but filled with placeholder assumptions. Ask it which portfolio companies have declining margins and it will explain what margin compression means – but it cannot tell you the answer, because it has no access to your numbers.

What MCP is

MCP – Model Context Protocol – is a standard developed by Anthropic that allows Claude to connect to external data sources in a structured, secure way. Think of it as a translation layer: MCP servers sit between Claude and your firm's systems, allowing the model to query, retrieve, and reason over your actual data without that data leaving your infrastructure or being used for model training.

The protocol is not a one-off API integration. It is a standardised framework, which means the same architecture that connects Claude to your CRM can also connect it to your deal pipeline platform, your portfolio reporting tools, your document management system, and your research databases. Each connection is an MCP server – a lightweight service that translates between Claude's query format and the specific data source's structure.

How it works in practice

Here is a concrete example. A partner at a PE firm sits down on Monday morning and asks Claude: "Which portfolio companies had revenue growth below 5% last quarter, and what were the primary drivers?" Without MCP, Claude would need the partner to copy-paste financial data into the conversation. With MCP, Claude queries the portfolio reporting system directly, retrieves the relevant data, identifies the companies meeting the criteria, and provides an analysis that references actual numbers from the firm's own reports.

The difference is not incremental. It is the difference between having a brilliant analyst who is blind and a brilliant analyst who can see your files.

Another example: a VP is screening a potential deal and wants to know if the firm has looked at similar companies before. She asks Claude: "Have we evaluated any industrial packaging businesses in the Southeast in the last two years?" Claude queries the CRM, pulls up relevant deal records, summarises the outcomes, and notes which team members were involved. This query takes seconds. Without MCP, the VP would spend 20 minutes searching through the CRM manually – or, more likely, would walk down the hall and ask someone who might remember.

What we connect

The specific data sources vary by firm, but the most common connections include CRM platforms – HubSpot, Salesforce, DealCloud, and Affinity are the four we implement most frequently. Each has its own data model and API structure, but the MCP layer abstracts that complexity so the end user experience is identical regardless of which CRM the firm uses.

Beyond CRM, we connect deal pipeline tools where active opportunities are tracked, portfolio reporting systems where financial and operational data from portfolio companies is aggregated, document management platforms where data room materials, IC memos, and research reports are stored, and external research databases that the firm subscribes to for market intelligence.

The key architectural principle is that each MCP server is purpose-built for its data source. The CRM server understands contact records, deal stages, and activity logs. The portfolio server understands financial statements, KPI dashboards, and reporting periods. The document server understands file structures, version histories, and content types. This specialisation is what allows Claude to ask the right questions of the right systems and return precise answers rather than approximations.

Security architecture

Data security is not an afterthought in this implementation – it is the foundation. MCP servers run within the firm's own infrastructure, or within a dedicated, isolated cloud environment that the firm controls. Data does not pass through Anthropic's servers for storage. The enterprise API agreement explicitly prohibits the use of customer data for model training.

Access controls mirror the firm's existing permission structure. If an associate does not have access to certain deal records in the CRM, Claude cannot access them on that associate's behalf. The MCP layer enforces the same role-based permissions that the underlying systems already apply. Every query is logged, creating an audit trail of what was accessed, by whom, and when.

For firms subject to regulatory requirements – which is most of our client base – the architecture is designed to satisfy compliance review. Data residency can be controlled (EU data stays in EU infrastructure, for example), and the entire system can be deployed in a private cloud environment with no public internet exposure if required.

Use cases we see most

Portfolio monitoring is the highest-frequency use case. Investment professionals ask Claude about portfolio company performance, trends, and anomalies on a daily basis. The queries range from simple ("What was Company X's Q4 revenue?") to complex ("Which portfolio companies have customer concentration above 30% and are also showing working capital deterioration?"). These questions previously required either manual spreadsheet analysis or a request to the portfolio operations team. Now they are answered in seconds.

Deal screening is the second most common use case. When a new opportunity enters the pipeline, Claude can pull comparable transactions from the firm's historical deal data, identify relevant research the team has previously produced, and flag any existing relationships with the target or its management team. This reduces the initial screening process from a half-day exercise to a 30-minute review.

Competitive intelligence is an emerging use case. By connecting to research databases and news monitoring tools, Claude can maintain a running brief on competitor activity, market dynamics, and regulatory developments within the sectors where the firm is most active. Partners receive a synthesis rather than a stack of articles.

Claude with MCP versus Claude without

The simplest way to understand the value of MCP is to consider the analogy of hiring a new analyst. On day one, that analyst is intelligent, well-trained, and eager – but they know nothing about your firm, your portfolio, your deals, or your process. They can produce generic work, but nothing that is grounded in your specific context. Give that same analyst six months of access to your systems, your historical memos, your portfolio data, and your deal records, and they become enormously more productive. They can answer questions with specificity, draft documents that match your standards, and identify patterns that only emerge from familiarity with your data.

MCP is what gives Claude that familiarity. Not through training on your data – the model itself does not change – but through live, structured access to the information it needs to produce work that is relevant to your firm, not just generically correct. The model's reasoning capabilities are the same either way. The difference is what it has to reason about.