# MCP server development: give AI secure access to your tools > We build custom MCP servers that let AI models work with your existing systems, with the permissions and logging a production system needs. We maintain an open-source MCP server ourselves. MCP (Model Context Protocol) servers are the bridge between AI models like Claude and your existing business tools. A server describes what a system can do as a set of tools, and any MCP-capable assistant or agent can then use them. The protocol is the easy part. The work is deciding which actions to expose, who is allowed to call them, and what happens when a model asks for something it should not get. We build MCP servers as stage two of our AI deployment program, and we publish one of our own under MIT. ## What we build - **Servers for your business systems.** Custom MCP servers that give AI direct, secure access to your CRM, ERP, databases, internal APIs and any other system in your stack. - **Tool design.** A small set of well-named tools with clear inputs and outputs, shaped around the tasks people actually want done rather than a mirror of your API. - **Authentication and permissions.** Read and write tools are split so they can be granted separately. Credentials are scoped to the smallest account that works, and anything that spends money or changes live data requires an explicit confirmation. - **Logging.** Every write is logged with the identity of the calling agent and its arguments, so you can see exactly what the AI did. - **Spec migration.** The July 2026 MCP spec drops sessions and the initialize handshake. We move existing servers to it and treat the change as a production release, not a dependency bump. - **Agents on top.** Once the tools exist, we build the skills and agents that use them to complete multi-step workflows. ## How we work 1. **Inventory the systems.** We list the tools your teams use and the tasks that cross them, and pick the few integrations that remove the most manual work. 2. **Design the tools before writing code.** We agree the tool list, the permissions and the failure behavior with you in writing. 3. **Test against real systems.** We run the server against your actual environment and check accuracy, latency, edge cases and failure modes. 4. **Deploy and document.** You get the server running in your environment, documentation, and a runbook. After handover, you own it. ## Where this comes from - [Promptify](https://matious.com/work/promptify.md): Several LLM providers sit behind one interface. We use the same principle for MCP: keep the differences between versions inside the integration layer, so the agents above it never have to know. - [TaxFigure](https://matious.com/work/taxfigure.md): Document intake turns tax forms into structured data that the research agents use as tools. Tool design for agents is the same discipline an MCP server needs. ## Frequently asked questions - **What is an MCP server?** A small service that exposes one of your systems to AI models as a set of tools, using the open Model Context Protocol. Any assistant or agent that speaks MCP can then use those tools. - **Which systems can you connect?** Anything with an API or a database: CRM, ERP, helpdesk, finance, document stores and internal services. If a system has no API, we look at what it can export. - **Is it safe to let a model act on our systems?** It is as safe as the permissions you give it. We split read and write tools so they can be granted separately, scope credentials narrowly, log every write, and require explicit confirmation before anything that spends money or changes live data. - **Have you published any MCP work?** Yes. We maintain the open-source Google Ad Manager MCP server, in Python under MIT. It manages campaigns, line items and creatives through natural language. - **Do you build on the new stateless spec?** Yes. New servers start on the July 2026 spec, and we migrate existing ones. Our article on the stateless spec covers what breaks and what gets easier to host. ## Further reading - [MCP goes stateless](https://matious.com/insights/mcp-2026-07-28-stateless-spec-guide.md) - [AI is now both attacker and target](https://matious.com/insights/ai-data-breach-pii-protection-llm-2026.md) - [December 1 on the Gemini API: a price rise and two forced model swaps](https://matious.com/insights/llm-price-change-model-switching-2026.md) ## Related - All services: https://matious.com/services.md - Case studies: https://matious.com/work.md - Start a project: https://matious.com/contact ## Canonical URL https://matious.com/services/mcp-server-development --- _Matious - AI engineering studio, founded 2017, Mohammedia, Morocco._ Contact: contact@matious.com | +212 661 790 436 | https://matious.com/contact Machine-readable: [llms.txt](https://matious.com/llms.txt) - [agent instructions](https://matious.com/agents.md) - [sitemap](https://matious.com/sitemap.xml) Every page is also available as Markdown: send `Accept: text/markdown`, or append `.md` to the URL.