Building V1s with intention: what InvestNext MCP does, what it doesn’t, and why both decisions matter

There is a specific kind of product discipline that does not get talked about enough: the discipline of what you decide not to build.

Features are easy to justify in isolation. Every capability sounds useful in a pitch. But the teams that build great software understand something harder to articulate: every addition has a cost, and sometimes the most important product decision is a deliberate subtraction.

That tension is where InvestNext MCP came from. And it is why I want to use this month’s development update to talk about what we built, what we intentionally left out for the present, and why I think both halves of that decision matter equally.

The scope question we sat with longest

When we started serious work on MCP, the scope question surfaced fast: how much should the connector be able to do?

The answer we kept coming back to was not about what was technically possible. It was about what we could honestly defend. InvestNext is the platform of record for LP data at the firms we work with. That is a specific kind of trust. Trust built over years of accurate records, clean audit trails, and a data model designed around the standards our clients are held to.

The question was not whether AI could be useful in that environment. It clearly can. The question was what kind of AI, doing what kinds of things, was appropriate to ship as a V1 in a category where the cost of a mistake is not a bad recommendation, but damaged LP confidence or a compliance exposure.

The answer we landed on was a read-only intelligence layer. A connector that gives GPs and IR teams natural-language access to their InvestNext data, inside the AI environment they already use, with no ability to write back to or modify anything in the platform.

That is not a limitation we’re rushing to eliminate: it is the product decision we are most confident in.

What read-only actually means, and why it matters

Read-only is not a technical constraint imposed by the protocol. MCP is fully capable of supporting write operations. We chose not to expose them…yet.

The reason is surface area.

Every action a system can take on a client’s behalf is a question that client has to answer for their LPs and compliance team. What did the AI do? On whose authority? What is the audit trail? A read-only intelligence layer has a clean answer to all three: it answered a question. That is the full extent of what it did. The failure mode is a wrong answer or no answer: not a record it wasn’t supposed to touch, a communication it wasn’t supposed to send, or a workflow it triggered without the right context.

In an industry where the trust chain runs from LP to GP to platform, the most responsible version of V1 is the one with the smallest audit surface. We will earn the right to expand that surface later, through Early Access feedback and real usage patterns that tell us where broader capabilities would genuinely serve our clients. But we want to balance that utility with where they would create more risk than value. 

We do that by building together with our clients toward a responsible, compliant, and safe agentic future, starting with our read-only MCP.

The data model decisions behind the connector

The protocol layer is only one part of the build. The harder product work was deciding exactly which InvestNext data to expose through the connector, and how to structure it for AI consumption.

A data model built for software UI and one built for natural-language AI access have different shapes. Dashboards render structured data visually; Claude synthesizes it into answers. The right data model for a chart is not necessarily the right one for a conversation.

We spent meaningful time on a few specific questions:

What does a GP actually ask AI before a meeting? Not in theory, but in practice. The answer shaped which investor record fields we prioritized, what summary structures we exposed, and how we handled data that exists across multiple records.

Where does permission-awareness break down under natural-language access? In the platform, user access controls are enforced at the UI level. With a connector, a question asked in natural language still has to respect the same access model. We built permission-awareness into the MCP layer itself, so the data is scoped identically to what that user can access in the platform directly.

What does “grounded in the data” actually require? Claude is a reasoning system. It synthesizes. That is the value. But synthesis that drifts from source data is worse than no synthesis at all in this context. 

What Early Access is teaching us

We launched the Early Access program earlier this month with a small cohort of sponsors who are already deep AI users. The goal was to build V1 alongside the people who would stress-test it hardest.

A few things the first weeks have already surfaced:

The questions GPs actually ask are more operational than we anticipated. Not “summarize my portfolio”. More like “which investors in Fund III haven’t been contacted in the last 60 days” or “what is the unfunded commitment position for my three largest LPs right now.” The specificity of real workflows is something you cannot design from the inside.

The moments of highest value are almost always time-constrained. Ten minutes before a call. Walking into a conference. The use case is rarely leisurely research, but someone who needs to arrive prepared and has very little runway to get there. That insight is shaping how we think about response structure and depth calibration going forward.

The trust moment is subtle but real. When an Early Access user gets an answer from Claude that they can trace directly back to their InvestNext records, AI stops feeling like a confident guesser and starts feeling like a reliable source. That is the moment we are designing toward.

What comes after read-only

I want to be honest about the roadmap in a way that does not overpromise.

The work in this category is going agentic. AI agents that can act on a sponsor’s data: executing routine tasks, drafting outreach, managing parts of the capital raise workflow with a GP in the loop. These are in the works from multiple platform providers. That is not a product vision we invented. It is where the technology is heading, and where our clients are already heading with it.

But agents require a level of grounding and trust that has to be earned before it can be extended. The source of truth has to be accurate. The permission model has to be airtight. The failure modes have to be understood. V1 of a read-only connector is how we build that foundation. It’s a prerequisite for doing agents responsibly in this category.

Early Access cohort partners are building that foundation with us. The questions they’re asking, the workflows they’re running, and the feedback they’re giving us on where the connector falls short are the input that shapes what we build next and in what order.

The product philosophy behind the decision

Matthew wrote last month about building something rigorous enough to honor the science of capital raising, while flexible enough to reflect the art of how deals actually get structured. I think about MCP the same way.

The science demands that LP data is accurate, auditable, and protected. The art is that GPs need to show up prepared for every conversation, at any moment, with information that is actually current. Those two things used to be in tension. You could have depth and accuracy, or you could have speed and accessibility, but not both without a significant team investment.

MCP is our path to hold both at the same time. Read-only now, agentic tomorrow, built together with our clients

The discipline of what we left out is what makes the rest of it sustainable and trustworthy for future iterations.

We’re still accepting applications for the Early Access cohort. If you’re already an InvestNext client, reach out to your account manager to learn more.

If you’re not an InvestNext client yet, but your firm has built real AI fluency into how you operate and you want your investment management data to meet you there, apply at investnext.com/mcp-early-access. I read every application.

— Kevin CEO, InvestNext