MCP: the standard that lets AI agents take action in your systems

AI agents can now reason through a fraud case, draft a client document, or summarize customer tickets. But reasoning isn't enough. To become true operational workers, AI agents need a safe, standardized way to act on your systems.

This has led to the development and rapid adoption of the Model Context Protocol as a way for agents to query systems. And, as we’ll see, if you run critical operations, especially in regulated industries, this also raises some very legitimate questions.

The problem with AI agents: reasoning without reach

Most operational stacks are complex. A fintech might run Forest Admin for backoffice operations, Resistant AI for document analysis, a third-party KYC provider for identity checks, Slack for internal coordination. Each system has its own API, its own terminology, its own authentication method.

An AI agent sitting on top of this type of fragmented stack needs to be able to reliably interact with it. One solution could be to write and maintain custom code for every single integration, every time.

That would be a bottleneck. And it's exactly why most "AI for operations" projects stall: not because the intelligence isn't there, but because the infrastructure to act on it isn't.

Enter MCP, the standard way to connect AI to tools

MCP, the Model Context Protocol, is an open standard, originally developed by Anthropic and now widely adopted across the industry.

In one sentence, MCP is a standardized way for an AI agent to call external tools, without writing custom code for each one.

Without MCP, every integration requires custom code to query an API. Your agent talks to Slack? Custom code. It calls your KYC provider? Custom code. It queries your ops database? Custom code. You're building a different bridge for every river.

MCP changes the architecture. Instead of writing a new integration each time, any tool that publishes an MCP Server becomes instantly accessible to any MCP-compatible AI agent. One standard. Interoperable by design.

A useful analogy: HTTP standardized how web browsers communicate with servers. It didn't matter which browser, which server, the protocol was shared. MCP is doing the same thing for AI agents and the operational tools they need to act on.

The two sides of MCP, to reach out or expose capabilities

MCP takes care of the two distinct roles found in any architecture: the server and client sides. Both are relevant for operational teams and their backoffice.

MCP Client: your system connects outward.

When your backoffice platform acts as an MCP Client, it can call any external MCP Server — a company registry API, a document verification service, a CRM, a Slack workspace — directly within your operational workflows. No custom code. Natural language as the integration interface.

MCP Server: your system opens inward.

When your backoffice platform acts as an MCP Server, it exposes itself — its data, its records, its workflow actions — to any AI agent that speaks MCP. A Dust agent, a Claude Desktop session, a custom internal AI tool: any of them can now query your operational layer, trigger actions, and get structured data back.

These are two different capabilities. But they serve the same strategic goal: making your operational system truly connected in the agentic era.

MCP isn't a niche trend, it’s the way forward to build agent platforms

If you've been tracking the AI tooling space, you'll have noticed: MCP adoption is accelerating fast.

Most major SaaS platforms now boast an MCP server to facilitate access. Slack, Monday, Attio or Intercom have shipped theirs. These aren't experimental bets, but product decisions by companies that understand where the market is heading and how their customer’s usage are evolving.

The shift is structural. AI agents are moving from being chat interfaces to being operational actors with a stronger degree of independence. For that to work safely at scale, there has to be a shared standard for how they interact with real systems. MCP is that standard.

For companies running regulated operations, like fintech or compliance-heavy industries, this way of opening systems to AI brings a whole new host of concerns. The stakes of an agent acting on the wrong data, with the wrong permissions, without an audit trail, are high. The MCP paradigm only works in these contexts if the platform it connects to brings governance, role-based access, and proper logging to the table.

Forest Admin is becoming the layer to access MCP securely 

Forest Admin is the first backoffice platform to implement both sides of MCP natively.

As an MCP Client: your Forest Admin workflows can now call external MCP Servers — pulling company registration data, triggering document analysis, querying third-party databases — without any custom development. Operations teams stop waiting for developers. Integrations become a configuration task, not an engineering project.

As an MCP Server: Forest Admin exposes your operational data and workflows to any MCP-compatible AI agent. A Dust agent can query your customer records. A Claude Desktop session can trigger an action. All within the permission model and audit trail already set up in your Forest Admin instance.

We use this ourselves. Our business team runs Dust agents — @CustomerOverview, @DealSnapshot — that connect directly to our own Forest Admin instance to access customer data and workflow runs. It's not a demo. It's how we work.

MCP Client and MCP Server are not two separate product bets. They're two sides of the same thesis: Forest Admin as the operational governance layer that agents can actually trust and act within.

What comes next

This article is the introduction. The next two will go deeper:

→ MCP Client: how your workflows connect to external tools without a single line of code — and what that means for ops teams currently blocked on developer dependencies.

→ MCP Server: how Forest Admin becoming an MCP Server turns it into the operational hub that your AI agents can pilot; with the right permissions, the right governance, and the right audit trail.

MCP isn't a feature update. It's the infrastructure layer for the agentic era. So if you're running regulated operations on Forest Admin, it's live now. Check it out and get in touch.