Skip to main content
News
Platform4 min read

Governing Every Tool Your Agents Can Call

One curated registry of connectors, and a check on every call before it reaches a system.

Governing Every Tool Your Agents Can Call

In From Cocoa to Chocolate, the AI Gateway governs what agents think with. The MCP Gateway governs what they act on. This piece walks through the factory’s equipment room: which machines exist, who may run each one, and what stops a machine that starts to misbehave.

From advisor to employee

A model that can only talk is an advisor. A model that can act is an employee, and an employee needs to know which machines it may operate.

MCP (Model Context Protocol) servers are the connectors agents use to act on real systems: to read records, create entries, search documents and trigger processes. They are what turns an answer into work. They are also where an agent with more access than its job needs can do real damage.

Without a gateway, any agent can reach any tool it has been given a connection to. Paid connectors bill per call with no cap. When something goes wrong there is no record to follow. The MCP Gateway keeps those connectors in one curated registry and decides who can call each one, how often and at what cost.

Every tool call is checked first

Before a call reaches an MCP server, the gateway checks three things in turn.

  1. Is the caller allowed this action?
  2. Does the connector’s budget allow it?
  3. Does the call rate allow it?
  4. The call reaches the server

Calls outside the rules never leave the gateway. A write that was never granted never reaches the system it was aimed at. The check happens before the action, not in a report afterwards.

A curated registry, by category

Every approved MCP server sits in one list, organised by category, showing its health, the number of actions it exposes and which agents use it. One look tells IT what agents can reach, and which agents reach it.

Adding a server is deliberate. Start from a template, or register any MCP server by its URL. The gateway checks the connection and loads the actions it offers. Servers can authenticate with an API key, OAuth or custom headers, or with each user’s own sign-in, so an agent acts with exactly the rights of the person behind it. Any server can be registered as read-only.

Nothing reaches the registry by accident. Every connector an agent can use is one someone chose to add.

Role-based access, down to single actions

The business defines which users, teams and agents can call which MCP servers, and within each server, which actions. Reading can be allowed while writing and deleting are denied.

Access is set per project and can be changed at any time, so capability is granted one tool at a time and withdrawn the same way. An agent starts with only what its job needs, and gains more only when someone decides it should.

Access is not all or nothing. An agent can be allowed to read a system without ever being able to change it.

Budgets and rate limits for paid connectors

Many connectors cost money per call. Each paid connector can be given a spend budget and a call rate per team or per agent, so one runaway process cannot run up a vendor bill, while every other team keeps working as normal.

Circuit breakers

When usage looks anomalous, such as call volume far above its usual rate, or when calls reach for sensitive data, the breaker trips and cuts the connector off before any damage spreads.

An agent caught in a loop keeps calling until someone notices. A circuit breaker is the part that notices first. Problems are contained first and diagnosed second.

Connectors for your stack, and your region

Connectors fall into three groups.

First-party products

Ready to connect: Visor, BuildPad, Workspace, Feedback and Chocolate Factory itself.

Enterprise systems

The systems of record that processes run on, such as CRM, SAP and Salesforce.

Local partners

Any partner service or system the business works with, including integrations not found anywhere else.

Together they let workflows run end to end, across the systems a business actually uses, under one set of rules.

From one server to every team

Rolling a connector out follows four steps.

  1. 1

    Register

    Put each MCP server in the registry once.

  2. 2

    Curate

    Group the tools each team needs into one set.

  3. 3

    Scope

    Decide who can call what, down to single actions.

  4. 4

    Watch

    See every call, what it returned and what it cost.

More in the MCP Gateway

  • Tool filteringEach agent sees only the tools it may use, so it cannot call what it cannot see.
  • Tool groupsChosen tools from several servers bundled into one set for a team or an agent. This is the Curate step.
  • Audit logEvery tool call recorded: which agent, which server and action, what came back and how long it took. This is the Watch step.
  • Cost per tool callWhat each call costs, next to the model spend behind it, by tool, server, team or agent.

What it changes

For the CIO and IT, every connector is in one registry. For risk, security and compliance, access is scoped by project and down to single actions, and circuit breakers stop anomalous or sensitive use. For finance, paid connectors carry their own budgets and rate limits, so one mistake does not become a large bill.

In chocolate terms, the food-safety inspector decides which machines each worker may run. The full feature set is on the MCP Gateway page.

The short version

  • Every tool call is checked for access, budget and call rate before it reaches a server.
  • Access is set per project for users, teams and agents, down to single actions.
  • Paid connectors carry their own budgets, and circuit breakers cut off anomalous or sensitive use.
  • Tool filtering, tool groups, an audit log and cost per tool call are part of it too.