Tools
Functions the model can call, described by a name, a description and a parameter schema (find a customer, read an order, create a ticket). This is the most widely used primitive.
AI integration · Model Context Protocol
MCP Server Development: Connect Your AI Agents to Your Tools and Data
Etixio builds MCP (Model Context Protocol) servers that let Claude, ChatGPT or your own AI agents use your tools and data in a controlled way: look up a record, search a document repository, create a ticket. We define the exposed tools, authentication, permissions and logging, then host and maintain the server in production.
Understanding the technology
A standard protocol between AI agents and your systems.
The Model Context Protocol is an open specification, released by Anthropic in late 2024 and handed over in late 2025 to the Agentic AI Foundation, under the Linux Foundation. It describes how an AI application (the MCP client) discovers and uses the capabilities exposed by an MCP server. Assistants such as Claude and ChatGPT, development environments and agent frameworks can all connect to an MCP server.
The benefit is that you build the connector once. Instead of developing a specific integration for each assistant or agent, you expose your business tools through an MCP server, with your own access rules. The server becomes the controlled gateway between the models and your information system.
Under the hood
What an MCP server exposes.
Resources
Content the client can read (documents, records, entries), identified by an address, to give the model context.
Prompts
Ready-to-use instruction templates offered to the user for recurring tasks.
Two transport modes
A local server communicates over standard input and output with the application that launches it; a remote server is exposed over HTTP and shared by several users. The July 2026 revision of the specification made the protocol core stateless, which brings hosting a remote server closer to hosting a standard web service.
Business use cases
What you can connect.
Your business tools for off-the-shelf assistants
Let your employees query your ERP, CRM or support tool from Claude or ChatGPT, with their own permissions.
Your internal agents
Give a custom-built agent standardized access to several systems, without rewriting every integration.
Your SaaS product
A software vendor can offer an MCP server to its customers so their AI assistants can work with the product’s data, within the limits of their subscription and permissions. It often complements AI features built into the product.
Your knowledge bases
Expose document search, for example a RAG engine, as a tool any MCP client can call.
The choices that matter
An MCP server is an access point to secure.
A model that calls your tools acts on the basis of text it has read, including content that may contain malicious instructions (prompt injection). Security must therefore be enforced by the server, not by the model.
Authentication
For a remote server, the specification relies on OAuth. We connect the server to your identity provider so that every call is tied to an identified user.
Least-privilege permissions
Each tool checks the user’s rights to the requested data. Read and write tools are kept separate, and sensitive operations require explicit confirmation.
Parameter validation
Inputs are checked server-side (types, ranges, formats), as with any exposed API, and results are limited to what is strictly necessary.
Logging and traceability
Every tool call is logged with the user, the client, the parameters and the result, for auditing and incident analysis.
Vetting third-party servers
We also help assess third-party MCP servers before they are approved in your organization, since a server installed without review can expose data or perform unexpected actions.
Hosting and operations
Hosting, monitoring and evolving the server.
A remote MCP server is a web service: we deploy it in your cloud or on your infrastructure, with its environments, monitoring, rate limits and alerts. Tool descriptions are carefully written and tested, because the model relies on them to pick the right tool and the right parameters.
We evaluate the server with the target clients (Claude, ChatGPT, your agents) on real scenarios, including ambiguous or unauthorized requests. The protocol evolves through dated revisions; we track these changes and the official SDKs (TypeScript, Python and other languages) to maintain compatibility. See our approach to LLMOps and AI evaluation.
In the field
From one-off connector to shared access point.
With Bobby, we connected a conversational assistant to our ERP data. This kind of connection can be exposed as an MCP server so several AI clients can use it, without multiplying integrations. For a medical AI agent, the same principle applies to document sources: permissions and traceability stay in the layer we build.
From work to deliverables
What we deliver.
- Inventory of the tools and resources to expose, with their access rules.
- An MCP server built, tested and documented, connected to your APIs and your identity provider.
- Evaluation scenarios with the target AI clients.
- Deployment, monitoring, logging and maintenance.
If an agent or MCP server prototype already exists, we can take it over and bring it to production. The MCP server is often part of a broader AI agents and chatbots or enterprise AI solutions project. It can be delivered as a fixed-price project or by a dedicated team that evolves it alongside your tools.
Frequently asked questions
MCP servers: your questions.
What is an MCP server?
It is a program that exposes tools, resources and prompts according to the Model Context Protocol. An MCP client (Claude, ChatGPT, a development environment or an agent) connects to it, discovers what is available and can call the tools on the user’s behalf.
How is MCP different from a regular API?
MCP does not replace your APIs; it usually builds on them. It standardizes how a model discovers and calls the available capabilities, with descriptions the model can read. A single server can therefore serve several assistants and agents.
Can an MCP server be used with both ChatGPT and Claude?
Yes, both can connect to remote MCP servers, under conditions that depend on the plan and your organization’s admin settings. We check these conditions for your target clients before design starts.
Is an MCP server secure?
It is if the server itself enforces authentication, permissions, parameter validation and logging. The protocol provides the framework but does not replace these controls. Sensitive actions should require confirmation.
Which language should an MCP server be built in?
Official SDKs exist for several languages, including TypeScript, Python, Go, Java, C# and PHP; the PHP SDK, still young, is maintained by the Symfony team with the PHP Foundation and plugs into Symfony through the MCP Bundle. We usually pick the language of your existing services so your teams can maintain it more easily.
Let’s connect your AI agents to your tools, under control.
Tell us which tools and data to expose, the users involved and the assistants you use. Together we will define the initial scope to explore.