Artificial intelligence
Model Context Protocol (MCP)
Publiée le October 19, 2025
Publiée le October 19, 2025
The Model Context Protocol, often abbreviated MCP, is an open-source standard designed to standardize the way AI models (especially large language models, LLMs) interact with external tools, data and services.
In other words, MCP acts as a “universal port” for AIs: rather like USB-C for electronic devices, it offers a unified way of connecting an AI model to databases, APIs, files, computational tools and so on.
The protocol was proposed by Anthropic in late 2024, and has since begun to be adopted in the AI ecosystem.
To sum up, MCP :
provides a common framework for making an AI model talk to other systems,
extends the context available to the model (not just what it “knows” via its training, but also what it can interrogate in real time),
aims to facilitate the integration of AI into operational environments (enterprise, business tools, autonomous agents).
Before the emergence of MCP, each integration of a language model with an external system (database, API, internal files, tools, etc.) required a custom connector. We ended up with an “M × N” integration problem: each model × each data source required specific code.
This multiplicity of ad-hoc interfaces made scalability, auditing and reuse difficult, and posed maintenance, security and governance challenges.
MCP responds to this dynamic by offering a standardized interface through which an AI (“client”) can discover, call and interact with services (“MCP servers”) without having to rework the whole connector each time.
This allows AI agent designers to focus more on business logic and “user intent” and less on technical plumbing.
MCP client: An AI model or AI agent that wishes to access external data, tools or services via MCP.
MCP server: A service that exposes resources (databases, files, APIs, calculations) according to the MCP specification. The client can access them in a standardized way.
Communication protocol: The protocol defines how requests are formulated, how context is passed on, how the model selects tools, and how responses are returned. For example, via JSON-RPC or function calls.
The IA model (client) queries the MCP server for “available tools” or contextual resources.
The customer can choose a tool or data source, and send a query in a standard format.
The MCP server processes the request, performs operations (e.g. data extraction, API call, file reading, calculation execution) and returns a result to the client.
The model uses this result as additional context to produce an answer, continue a flow of reasoning or trigger other tools.
Suppose an AI agent is tasked with helping a developer modify a code repository:
The model queries an MCP server, which exposes a GitHub repository.
It asks “show me file X” or “analyze this function”.
The MCP server returns the content.
The model proposes a modification (pull request) and the MCP server can perform the action (“create PR”).
This type of use was demonstrated by Anthropic when it launched MCP.
Reduction of integration-specific code.
Decoupling of AI model and business connectors.
Dynamic tool discovery (the agent can decide which tool to call up).
Better traceability and governance of data access.
Here are the main benefits of the MCP protocol:
With MCP, developers don’t have to write specific connectors for each new model or data source. A single interface is all that’s needed, reducing development time and complexity.
Language models are no longer limited to what they have learned during training: they can query data in real time, access workflows and manipulate files, making them much more useful in an enterprise context.
Client-server architecture means that new MCP servers can be added (e.g. access to a new database or business tool) without modifying the IA model itself. This favors a “plug and play” architecture.
In a corporate context, it’s vital to know who is accessing what, in what context, and what data is being used. MCP facilitates auditing, selective data filtering and data localization (an MCP server can be deployed on site).
Since MCP is an open standard, different AI models, connectors and tools can work together consistently. This level of interoperability is rare until now.
Like any emerging technology, MCP also presents challenges that are important to be aware of.
Giving an AI model real-time access to files, APIs or databases poses risks: prompt injections, data exfiltration, malicious use. A study of vulnerabilities on numerous MCP servers has shown that certain protocol-specific flaws do exist.
It’s essential to implement precise controls: which tool can be invoked, which data can be exposed, which actions can be taken. Without this, you leave yourself open to abuse or data leakage.
Although promising, MCP is still in its infancy: specifications, SDKs and the ecosystem of tools are still being built. Adoption is not yet universal, which may limit the benefits in certain contexts.
When an AI model links up several tools (“multi-tool” workflow), keeping track of context, dependencies and errors becomes complex. MCP facilitates this, but the challenge remains.
| Criteria | Ad hoc integrations | Model Context Protocol (MCP) |
|---|---|---|
| Number of connectors to develop | New connector for each source/model | One standard protocol for many sources |
| Maintenance & upgradability | Complex to maintain | Modular, reusable, scalable |
| Model contextualization | Limited, often static | Real-time dynamic access to data/tools |
| Governance & audit | Often cobbled together | Integrated via protocol, better traceability |
| Interoperability | Low – between models/tools | High – model, tool, connector work together |
| Security & control | Varies according to implementation | Requires rigor but protocol designed for it |
In this way, MCP represents a significant advance over traditional methods of integrating AI models into real systems.
Here are a few examples of how MCP can be deployed:
Code development agent: An AI assistant integrated into an IDE can, via MCP, access the code repository, analyze existing functions, propose modifications and carry out pull requests.
Internal enterprise chatbot: An AI agent that queries the company’s knowledge base, internal files, CRM, ERP, via on-site MCP servers, to respond to employees quickly with up-to-date data.
Workstation-based Personal Assistant: The template accesses your personal files (with permission), reads documents, prepares a summary, schedules tasks or automates system actions – via a local MCP server.
Multi-tool agentic flow: A model which, for a single user request, links document search → calculator API → database → file update, all coordinated via MCP.
These uses show that MCP isn’t just a technical curiosity: it’s a catalyst for more useful, contextualized and powerful AI.
It accelerates the integration of AI into companies. Instead of redoing the plumbing for each use case, we have a standard.
It enhances the user experience: AI can act not only on what it “knows”, but also on what it can do and access.
It promotes interoperability between different suppliers of AI models, tools and platforms. This reduces lock-in and creates a more open market.
It strengthens governance and compliance: in a context where data access, traceability and accountability count for a lot, having a standard protocol is an advantage.
It paves the way for an “agentic” AI architecture where models, tools and workflows can be orchestrated in a modular way, a step towards autonomous AI in a controlled setting.
Broad adoption and maturity: For MCP to truly become a standard, it needs to be adopted by a large number of players (AI models, tool providers, enterprises) with robust implementations.
Security & trust: As already mentioned, enabling an AI model to perform actions or access sensitive data requires strong guarantees (authentication, permissions, logging, sandboxing).
Performance & latency: Accessing external sources can introduce latency. We need to ensure that the end-user experience remains fluid.
Interoperability between versions: As with any young standard, there may be divergent implementations, incompatible versions, or fragmentation.
Ethical governance: We must ensure that AI does not abuse its ability to access/act, that user consent is clear, and that transparency is maintained.
Connector maintenance: Although MCP reduces connector overload, the MCP servers themselves need to be maintained, secured and audited. The study shows that many open-source servers have “code smells” or flaws.
The Model Context Protocol (MCP) marks a turning point in the way artificial intelligence systems are designed and deployed. By providing an open standard for connecting AI models to external services, data and tools, it makes these systems more powerful, contextual, integrated and governable.
For companies looking to exploit language models operationally – and not just experimentally – MCP paves the way for :
more modular workflows,
reduced time-to-production,
strengthened governance,
improved interoperability.
Of course, the protocol is not without its challenges: security, governance and ecosystem maturity remain points to watch. But few recent technologies combine so much promise to take AI from the lab to the enterprise.
If you’re planning to integrate AI agents into your organization, MCP deserves to be at the heart of your architectural thinking.