MCP Connectors, explained: What they are, how they work, and who should care
By Shivani Kaul, Head of Marketing, DwellFi
An AI agent that can read your fund accounting system but cannot write back to it is a demo, not an operation.
MCP connectors close that gap, and the gap is where most back office automation projects in private markets stall. The agent summarises a capital account statement beautifully, then a human retypes the result into Investran. A connector gives the agent the same access to a system of record that a named person has, under the same permissions and the same audit trail. Every connector DwellFi builds ships as an MCP server for exactly that reason.
What Are MCP Connectors?
An MCP connector is an implementation of the Model Context Protocol, an open standard that defines how AI systems call external tools and data. The connector sits in front of one system, exposes that system's operations as tools an agent can call, and handles authentication, permissions, and logging on every call.
MCP in plain language
Think of it as a driver. Your operating system does not know how to talk to a specific printer; the driver does, and it presents a standard interface upward. MCP does the same thing for AI. The agent learns one calling convention. Each connector translates it into whatever dialect the target system speaks.
Why this matters for private markets data
Fund data does not live anywhere. It lives in a general ledger, a sub-ledger, a custodian feed, an investor portal, and a folder of SFTP drops that someone set up in 2019. Each one has its own API, its own authentication, and its own idea of what a valid date looks like. Agents built against that fragmentation spend most of their reasoning budget on plumbing. MCP collapses the dialects into one interface, so the agent can spend its attention on the reconciliation instead of the connection.
How DwellFi's MCP Connectors Work
Nothing gets installed in your estate. One outbound link reaches back.
Extract, map, validate, deliver
Four stages. The connector reads the source, applies mapping rules, holds anything that fails validation, and writes to the target. Every row carries its own trace: where it came from, which rule version touched it, what the result was, and when. Give us an identifier and we return its full path between systems.
One outbound link, one dedicated tenant
The connection is a single outbound TLS 443 link from your network. No inbound ports. No VPN. No software on your servers. That matters because it turns a six-week security review into a conversation your network team can approve in an afternoon.
Your environment is single-tenant, with its own database and storage. Your data is not pooled with another client's. Batches in the millions of rows run as one vectorised pass rather than row by row, so a growing fund count does not change the shape of the job.
Key Use Cases: NAV Reconciliation, LP Reporting, and Capital Calls
NAV reconciliation
Agents reconcile across custodians, the accounting system, and market data feeds continuously rather than at quarter end. Breaks surface as a worklist grouped by cause, not as a 4,000-row spreadsheet someone opens on a Friday. The connector is what makes it continuous: the data arrives as it changes rather than as a nightly export, so a pricing break found on Tuesday is a Tuesday problem instead of a quarter-end one.
LP reporting
Investor statements and performance views generate from live fund data, formatted to each LP's preference. The work that disappears is not the writing. It is the aggregation, the copy-paste between three systems, and the version-control argument about which figure is current.
Capital calls
Call notices get parsed, investor data validated against the registry, and notices prepared without manual re-keying. Because the connector is bi-directional, the resulting entries post back into the system of record rather than sitting in a report nobody can action.
The pattern holds across all three. The agent does the reading and the drafting. The connector does the moving and the proving. Neither is useful without the other, which is why connector quality decides how far an agent programme actually gets.
Supported Systems and Integrations
DwellFi's catalog covers 1,526 pre-built connectors. The ones that matter for fund operations: fund accounting systems, bank feeds, SFTP, and cloud storage, with direct bi-directional sync and no schema changes on your side.
NetSuite by Oracle runs through NetSuite's own MCP server rather than a proprietary bridge. Worth saying plainly, because it answers the vendor-lock question before anyone has to ask it.
Beyond the fund stack, the catalog reaches the systems the work actually touches: email and document stores where capital call notices arrive, databases and warehouses where the reporting layer lives, and the payment systems on the other end of a distribution. An agent that can read a call notice in Outlook, validate it against the registry, and post the entry is doing one job, not three handoffs.
What Happens When a Row Fails
This is the part most integration content skips, and it is the part that decides whether anyone trusts the pipeline.
A bad row does not stop the batch and does not vanish into a log file. The good records commit on time. The failures go to a quarantine table with the source payload and the reason attached.
Then four steps run on the held rows only. The agent groups the failures by cause and works out what changed. It drafts the mapping rule that would clear them. The draft runs against the held rows and recent history, so the recommendation arrives with its test result already attached. A named person approves, and only the held rows re-run.
Anything touching money or entity matching waits for a person. Always. Triage is a recommendation, never a silent edit, and both the proposal and the approval are recorded against the rule version. Switch the agent off and you still have the queue.
MCP Connectors vs. Traditional API Integration
ㅤ | Traditional API integration | MCP connector |
Built for | An application you wrote | An agent that reasons about the call |
Permissions | Service account, often broad | Same permission model as a person |
Audit trail | Application logs, if you built them | Per-row trace with rule version |
Cost basis | Per call, per record, or per seat | Fixed monthly fee per connector |
The cost line is the one that changes behaviour. One fixed fee per connector, with nothing charged per record, fund, client, or API call. Adding funds or data points to a live connector costs nothing, which means nobody has to ration queries to protect a budget.
Frequently Asked Questions about MCP Connectors
What is an MCP connector?
An MCP connector is software that exposes one system's operations as tools an AI agent can call, using the Model Context Protocol. It handles authentication, permissions and logging so the agent works under the same controls as a person.
Is MCP the same as an API?
No. An API is built for an application you wrote. An MCP connector is built for an agent that decides which call to make, so it carries tool descriptions, permission scoping and per-call audit logging the API does not provide.
Do MCP connectors require software installed on our servers?
No. DwellFi connectors use a single outbound TLS 443 link from your network. No inbound ports, no VPN, and nothing installed in your estate.
Can an MCP connector write data back into our fund accounting system?
Yes. DwellFi connectors are bi-directional, reading from and writing to accounting systems with no schema changes on your side.
How to Get Started
Pick the reconciliation you run most often, not the one you find hardest. Frequency compounds; complexity just makes the first build slow. Prove one connector end to end, watch the exception queue for a month, then widen across the rest of the fund administration stack.
Where a system publishes an API description, the connector tools are generated from that description rather than hand-coded. That is why a new connector takes weeks rather than quarters.
Book a technical discovery session and bring one workflow. We will show you the exception queue, which is the honest part of any integration demo.