# What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open protocol that standardises how AI applications connect to external data sources and tools. In finance it matters because it can put system functions within an agent's reach, so consent, authorisation and logging must sit outside the model.

- Source: https://finray.tech/intelligence/glossary/model-context-protocol/
- Published: 2026-11-21
- Updated: 2026-11-21
- Publisher: Finray Technologies Ltd, Cyprus Companies Registry HE 445903

---

## What is the Model Context Protocol?

The Model Context Protocol (MCP) is an open protocol that standardises how AI applications connect to external data sources and tools. Its specification describes it as a way to share contextual information with language models, expose tools and capabilities to AI systems, and build composable integrations and workflows. For a financial institution it matters because a tool exposed through MCP can put a system function, such as reading a ledger or submitting a payment, within an AI agent's reach.

MCP is vendor-neutral. Anthropic introduced MCP in November 2024; since December 2025 it has been a project of the Agentic AI Foundation under the Linux Foundation. This entry explains the protocol and the controls that belong around it. It is not a statement that any particular vendor offers an MCP integration.

## How does MCP work?

The specification describes three roles that communicate with JSON-RPC 2.0 messages:

- **Hosts** are the AI applications that initiate connections.
- **Clients** are connectors inside the host.
- **Servers** are services that provide context and capabilities.

Servers can offer three kinds of feature: resources (context and data), prompts (templated messages and workflows) and tools (functions for the model to execute). Clients can offer elicitation, which lets a server request more information from the user. The specification also defines optional extensions, which are opt-in and negotiated through declared capabilities. Details such as protocol versions change over time, so read the current specification rather than relying on a summary.

In practice, a host connects to a server, discovers what tools it offers, and lets the model request tool calls. The host decides whether and how a call is executed.

## Why does MCP matter in finance?

Before protocols like MCP, each AI integration with a system needed custom work. A standard way to expose tools lowers the cost of connecting an agent to many systems. That is useful, and it also widens the set of things an agent can attempt. For a payment or accounting function, a tool call may be an instruction to change money or records.

MCP is a general integration protocol, not a payments protocol, so it defines no payment semantics, limits or authorisation model for moving money. Those have to come from the systems behind the tools.

## What does the specification say about security?

The specification has a Security and Trust and Safety section. Its key principles are:

- **User consent and control.** Users must explicitly consent to and understand all data access and operations, and retain control over what is shared and what actions are taken.
- **Data privacy.** Hosts must obtain explicit consent before exposing user data to servers.
- **Tool safety.** Tools represent arbitrary code execution and must be treated with appropriate caution. Descriptions of tool behaviour, such as annotations, should be considered untrusted unless obtained from a trusted server, and hosts must obtain explicit user consent before invoking any tool.

The specification adds that MCP cannot enforce these principles at the protocol level, and that implementors should build robust consent and authorisation flows, document security implications and implement appropriate access controls. For a financial institution this is the central point: the protocol carries the call, and the institution's systems have to decide whether to honour it.

## What risks does MCP add in a regulated setting?

| Risk | What can go wrong | Control |
|---|---|---|
| Over-broad tool scope | An agent can call functions beyond its task | Expose narrow tools; scope credentials per agent and task |
| Untrusted tool descriptions | A server's description misleads the model | Use only trusted servers; review descriptions; treat annotations as untrusted |
| Injection through content | Text in a document or page tells the agent to call a tool | Authorisation in the target system; approval for consequential calls |
| Write operations repeated | A retry executes a payment or posting twice | Idempotency keys on write tools |
| Credential misuse | A token is used outside its purpose | Short-lived, revocable, scoped credentials |
| Missing audit trail | Nobody can show what the agent called | Append-only log of calls, arguments, results and approvals |
| Third-party dependency | The model or server provider changes behaviour | Inventory and monitoring of AI components; contractual and technical controls |

## What do supervisors expect?

FINMA's Guidance 08/2024 expects supervised institutions to identify, assess, manage and monitor risks from the use of AI, and lists operational and model risks, data risks, IT and cyber risks, growing dependence on third parties, and legal and reputational risks among its focus areas. Connecting a model to institutional systems touches most of those areas at once.

The EU AI Act, for systems classified as high-risk, requires that they can be effectively overseen by natural persons while in use (Article 14) and that they technically allow automatic recording of events (Article 12). Whether a given agent setup is high-risk depends on use and classification, and application dates are staged.

## What is a worked example?

The numbers are synthetic. The example illustrates the control pattern; it is not a description of any one product's features. A bank exposes two tools through an MCP server for an internal operations agent: "read account balance" and "submit payment instruction". The credentials given to the agent allow the first freely and the second only up to CHF 2,000 per instruction, with anything above routed to a human approver.

The agent reads a balance of CHF 48,200 and proposes a payment of CHF 1,750 to an approved payee. The payment system, not the agent, checks the limit and the payee, and executes it with an idempotency key. The agent then proposes CHF 9,500. The payment system holds the instruction and raises an approval task. A message in a document the agent read told it to "send CHF 9,500 immediately"; the host's consent flow and the payment system's limit both prevent execution without a person.

The log shows the tool calls, the arguments, the limit check, the approval request and the outcome.

## What belongs outside the model?

- **Authorisation.** Enforced by the system behind each tool, using the caller's identity and scope.
- **Limits and approvals.** Evaluated by code and independent of the model's output.
- **Consent.** The host obtains user consent as the specification requires; the institution adds its own policy for consequential tools.
- **Idempotency.** Write tools accept idempotency keys.
- **Logging.** Every call, argument and result is recorded in an append-only log.
- **Server vetting.** Servers are inventoried and trusted deliberately, with descriptions reviewed.

## How does Finray relate to MCP?

Finray does not claim agent integrations through MCP for [Corebanq](/platforms/corebanq/) or [Ordinis](/platforms/ordinis/). Corebanq Accounting offers chat access from Claude or ChatGPT. Questions are read-only. Documents arrive as drafts; a person approves before anything posts. Chat access is off by default and activated per company. The controls described above are properties an institution should look for in any system it considers exposing to agents. On the platform pages, the relevant existing foundations are role-based access control in Corebanq where every action carries the actor and the policy version, approval routing in Ordinis, and second-person approval before posting, on by default, in [Corebanq Accounting](/platforms/corebanq-accounting/), whose AI assistant is off by default, runs on a model the customer hosts locally and has no posting authority.

Related reading: [agentic payments](/intelligence/glossary/agentic-payments/), [agent mandate](/intelligence/glossary/agent-mandate/), [human-in-the-loop approval](/intelligence/glossary/human-in-the-loop-approval/), [agentic commerce](/intelligence/glossary/agentic-commerce/).

## Frequently asked questions

### What is the Model Context Protocol?

MCP is an open protocol that standardises how AI applications connect to external data sources and tools. It uses JSON-RPC messages between hosts (the AI application), clients (connectors inside the host) and servers (services that provide context and capabilities).

### Is MCP a payments protocol?

No. It is a general integration protocol. A payment or accounting function can be exposed as an MCP tool, but MCP itself defines no payment semantics, limits or authorisation model for money movement.

### What does the MCP specification say about safety?

It states that users must explicitly consent to data access and operations, that hosts must obtain explicit user consent before invoking any tool, and that descriptions of tool behaviour should be considered untrusted unless obtained from a trusted server. It also says MCP cannot enforce these principles at the protocol level, so implementors must build the consent and authorisation flows.

### Does Finray offer an MCP server?

Finray does not claim agent integrations through MCP. This entry is a vendor-neutral explanation of the protocol and the controls a financial institution should apply if it exposes systems through it.

### What controls should surround an MCP tool that touches money or books?

Authorisation and limits enforced in the system behind the tool, not in the model; scoped credentials; human approval for consequential actions; idempotency on write operations; and an append-only record of every call and its result.

## Sources

- [Linux Foundation, Formation of the Agentic AI Foundation (9 December 2025)](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation)
- [Model Context Protocol specification](https://modelcontextprotocol.io/specification/latest)
- [FINMA Guidance 08/2024, Governance and risk management when using artificial intelligence](https://www.finma.ch/en/news/2024/12/20241218-mm-finma-am-08-24/)
- [Regulation (EU) 2024/1689 (AI Act), EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689)
