Table of Contents

• The short answer

• Why MCP matters for support

• How it works in Fini AI

• Safety: what stops the agent doing the wrong thing

• Tool catalog states

• MCP or REST Actions?

• A rollout that keeps risk low

• Frequently Asked Questions

The short answer

An MCP customer support agent connects to your Model Context Protocol server and uses the tools it exposes. In Fini AI, read tools become Attributes that give the agent customer context, such as order status or plan. Write tools become Actions that run at Tool nodes inside the Intent Rules you design, such as issuing a refund or changing an address. Tool schemas sync automatically, so there is no field mapping to maintain.

Have an MCP server already? Book a demo.

Why MCP matters for support

Support answers depend on live data: where the order is, which plan the customer has, whether a refund is allowed. Without that, an AI agent can only quote the help center and hand everything else to a person.

The usual fix is one custom integration per system. Each one needs field mapping, and each breaks when the system changes. MCP gives one standard interface: your team runs one server, exposes the tools it wants, and keeps control of the permissions behind them.

How it works in Fini AI

1. Connect your MCP server. Share your MCP server details; the Fini team connects it and runs the first sync. Fini AI then discovers the tools it exposes.

2. Read tools become Attributes. Order status, plan, account flags and similar context flow into the conversation.

3. Write tools become Actions. They run at Tool nodes in your Intent Rules, so a refund or account change happens only in the workflow you designed.

4. Schemas stay in sync. Fini AI listens for tool list change notifications and also polls, so new or changed tools appear without remapping fields.

How Fini AI does this: one MCP connection turns your read tools into customer context and your write tools into Actions inside Intent Rules, next to the native integrations. See Fini AI integrations.

Safety: what stops the agent doing the wrong thing

Tool hints set safe defaults

Fini AI reads the MCP hints on each tool (readOnlyHint, destructiveHint, idempotentHint) and uses them to set safe defaults. Any tool without readOnlyHint is treated as a write tool, and a tool marked destructive starts with tighter defaults than a read-only lookup.

You set the policy

You choose which tools the agent may call. Nothing on your server is used until you enable it.

Actions run inside your workflow

A write tool can't be called freely in conversation. It runs at a Tool node in an Intent Rule, after the Check nodes you put before the Tool node, such as identity verification or a refund limit.

Secrets stay out of the conversation

Secrets and internal IDs are hidden from the model and redacted from logs. For write calls, Fini AI injects an idempotency_key so a retried call doesn't run twice. In the chat widget, a signed widget token carries the customer's identity, so the agent never relies on what the customer types about who they are.

Every step is traced

Each conversation's AI Steps trace shows which tools ran, with what inputs, and which rule triggered them, so you can audit any action after the fact.

Tool catalog states

MCP or REST Actions?

Fini AI also calls REST APIs through Actions and connects natively to helpdesks, commerce tools and knowledge sources (see integrations). Use MCP when your internal systems already expose an MCP server, or when you want one interface for many tools. Use REST Actions for a single endpoint you already have.

A rollout that keeps risk low

1. Expose read tools first and enable them as Attributes.

2. Check answers on real conversations using the AI Steps trace.

3. Add one write tool behind an Intent Rule, with Check nodes before the Tool node.

4. Widen the rule only after the action has run cleanly on real requests.

The rest of the agent, from knowledge to Intent Rules, lives on the Fini AI platform.

See how Fini AI handles MCP. Book a demo.

Frequently Asked Questions

Can an AI customer support agent use MCP?

Yes. Fini AI connects to your MCP server, uses read tools for customer context and turns write tools into Actions that run at Tool nodes inside your Intent Rules.

Do we need to map fields to connect an MCP server?

No. Tool schemas sync continuously, so there is no field mapping. Share your server details and the Fini team connects it and runs the first sync.

Can the agent call any tool on our MCP server?

No. It can only call tools you enable, and write tools run only after the Check nodes you place before the Tool node.

How does an MCP support agent avoid destructive actions?

Fini AI reads each tool's hints to set safe defaults, treats any tool without readOnlyHint as a write tool, and runs write tools only inside Intent Rules you design.

What happens when our MCP server changes?

New and changed tools are picked up automatically through list change notifications and polling. A tool that stops responding is marked Degraded and the conversation routes to Fallback.