The Model Context Protocol (MCP) is an open standard that lets AI agents call external tools and read data through small adapter programs called MCP servers. A coding agent such as Claude Code, Codex, Cursor or Copilot can use MCP servers to read issues, query a database, browse documentation or control a browser. They are powerful and risky in equal measure, so teams should allow only vetted servers with minimal permissions.
On this page
A coding agent is only as useful as the context it can reach. Out of the box it sees your files and can run commands. Model Context Protocol servers extend that reach to issue trackers, databases, design files, monitoring tools and internal APIs. Anthropic introduced MCP as an open protocol in late 2024, and it has since been adopted by most major agent tools, including Claude Code, OpenAI Codex, Cursor and GitHub Copilot. This article explains how MCP works, which servers are worth considering for development teams and how to use them without creating new security holes.
How MCP works
MCP defines a common language between an AI application, called the client or host, and an MCP server that exposes capabilities. A server offers three kinds of things: tools, which are actions the model can call such as creating a ticket or running a query; resources, which are data the client can read such as a document or a schema; and prompts, which are reusable templates. Servers run either locally, started by the agent as a process and communicating over standard input and output, or remotely over HTTP, often with OAuth for authentication. Because the protocol is shared, one server works with every compatible agent, which is the main reason MCP spread so quickly. The official specification and SDKs are linked from modelcontextprotocol.io.
Useful MCP servers for development teams
The most valuable servers connect the agent to information it would otherwise ask you for. Issue trackers and source hosting, for example GitHub, GitLab, Jira or Linear servers, let the agent read the ticket it is working on and link its pull request. Documentation servers fetch current library docs, which reduces outdated API usage. Database servers let the agent inspect schemas and run read-only queries in development. Browser automation servers such as Playwright-based ones let the agent open the app, click through a flow and check the result. Observability servers for error tracking or logs help with debugging production issues. Design tool servers can pass component specifications to the agent for front-end work.
Setting up MCP in your agent
- Start with one concrete problem, such as the agent not knowing ticket details, and pick a single server that solves it.
- Prefer official servers maintained by the vendor of the tool you connect, and read their source or documentation for what each tool does.
- Configure the server at project level so the whole team uses the same setup; Claude Code, Codex and Cursor each support project configuration files for MCP.
- Give the server the narrowest credentials possible: read-only database users, scoped API tokens and no production access.
- Test it with a few prompts and check in your agent's logs which tools it actually calls.
- Document the server, its purpose and its owner in your context file, and review the list quarterly.
Security risks you must manage
MCP servers execute code and hold credentials, so they belong in your threat model. The main risks are prompt injection, where content returned by a server, such as an issue description or web page, contains instructions that hijack the agent; excessive permissions, where a server can write or delete far more than needed; untrusted servers installed from unknown sources, which can exfiltrate code or secrets; and data leakage, where sensitive data from a database ends up in prompts sent to a model provider. Mitigate these by allowlisting servers, requiring approval for write actions, isolating agent sessions and never connecting production systems with personal data to a development agent. Our security checklist for AI coding agents lists further controls.
MCP in larger organizations
Once several teams use agents, MCP becomes an architecture question. A central list of approved servers, a registry of internal servers that wrap company APIs, and logging of tool calls make the setup auditable. Enterprises that want this governance for AI use beyond coding, with compliance features and custom APIs, can look at platforms such as KalaidoCreate, which focuses on enterprises and the public sector. For data that must not reach a model at all, anonymize it before it enters any agent workflow.
Common mistakes
Teams often install too many servers at once. Each server adds tool descriptions to the agent's context, which can confuse tool selection and consume tokens. Another mistake is building a custom server before checking whether a simple command-line tool the agent can already run would do. Finally, remote servers with OAuth are convenient, but tokens stored on developer machines need the same care as any other credential. Where MCP shines is in combination with clear tasks, as described in multi-agent development workflows.
Learning MCP hands-on
MCP is easiest to understand by connecting one real server to one real task. The neooutput agentic coding workshops include MCP setup and security exercises on Claude Code, Codex and Cursor, so your team leaves with a configuration it can use the next day. At OnlineAppCreator, where we build and launch digital products with AI Code Engineers, a small set of well-chosen MCP servers is part of every project setup.
Frequently asked questions
Is MCP only for Claude?
No. Anthropic created the protocol, but it is open and supported by many clients, including OpenAI Codex, Cursor, GitHub Copilot and other agent tools.
Do I need to write my own MCP server?
Usually not at first. Official servers exist for many common tools. Write your own when you need to expose an internal API or database to agents in a controlled way.
Can an MCP server access my production database?
It can if you give it credentials, which is why you should not. Use read-only access to development or anonymized data, and keep production systems out of agent sessions.
Do MCP servers increase costs?
Indirectly, yes. Tool descriptions and returned data add tokens to every request, so only enable the servers a project actually needs.