The new integration everyone will expect
For two decades, “integrations” meant a REST API and a Zapier app. That’s changing. Your users increasingly work through AI assistants. They ask Claude to update the CRM, ChatGPT to pull last month’s numbers, an agent to file the ticket. The Model Context Protocol has become the standard way for those assistants to connect to products, and search interest in “MCP server” climbed steeply through 2025 and 2026.
Few engineers have built production MCP servers with proper authentication and security, so competition for this work is still light. The products that become easy for AI to use now will be the defaults agents reach for later.
What separates a production MCP server from a weekend demo
Tool design. The most common mistake is generating one tool per API endpoint. Faced with sixty similar tools, models pick the wrong one or chain five calls where one should do. I design a small set of task-level tools with precise descriptions and constrained, typed inputs. Then I test which ones models actually choose correctly.
Authentication and permissions. A remote server needs OAuth tied to your real user accounts, scoped tokens, and permission checks in every tool. The AI acts as the user, never with more access than the user has.
Safety against agent-specific risks. Agents read untrusted content: emails, web pages, tickets. That content can contain instructions meant to hijack them. Confirmation steps for destructive actions, sensible defaults, rate limits and audit logs keep a confused agent from becoming an incident.
Evaluation. Unit tests prove the server works. Agent evaluations prove the models use it correctly. I run realistic task prompts through several AI clients and track success rates as the tools evolve.
A typical engagement
Most MCP projects follow the same arc.
- Week 1: design. I sit with your product and engineering leads and list the five to fifteen jobs customers most often want done: “find overdue invoices and send reminders”, “summarise this account’s activity”, “create a project from this brief”. Each becomes a candidate tool, with clear inputs, outputs and permission requirements.
- Weeks 2–3: build. The server is implemented against your existing API or database layer, with OAuth wired to your identity provider. Every tool is unit-tested, and destructive tools get confirmation flows.
- Week 4: evaluate and harden. Realistic prompts are run through several AI clients. I measure how often models pick the right tool and fill its arguments correctly, then rewrite descriptions and schemas where they stumble.
- Launch. Deployment, monitoring, public docs and directory submissions follow.
Industry pricing for production MCP servers typically sits between about $10,000 and $30,000, depending on how many tools you expose and how complex authentication and permissions are. The fixed-price build above covers a focused first version. Most products get the majority of the value from a well-chosen dozen tools, and you can expand from there.
Systems experience where it counts
An MCP server is a networked service under load, called by unpredictable clients. I’ve built and run exactly that kind of system: an end-to-end game network optimizer. It has signed session tokens checked offline, replay protection, a NestJS API with licensing and audit logs, and CI-deployed servers. The same habits apply here: authentication done properly, defensive input handling, and observability from day one.
You don’t need to rebuild anything to get started. An MCP server sits in front of the API you already have, translating agent-friendly tools into your existing endpoints. Your core product stays exactly as it is.
Pairs well with
- A RAG assistant, so agents can both read your documentation and act in your product.
- Workflow automation that uses your MCP tools inside internal business processes.