The problem MCP solves
A language model writes and reasons well, but knows nothing about your records. It can't look up a client, schedule a meeting or log hours. For that it has to talk to other software.
Before MCP, every AI app had to build its own integration for every program. MCP turns that into one agreement: a program builds an MCP server once, and every assistant that speaks MCP can use it.
Think of a power socket: the appliance and the grid don't need to know each other, as long as they use the same plug.
How MCP works
There are two sides. The server sits with the program that holds the data, a time tracker for example. The client sits in the AI assistant: Claude Desktop, Claude Code, ChatGPT, Cursor, Codex or VS Code.
The server describes the tools it offers, each with a name, a plain-language explanation and the fields it expects. The assistant reads those descriptions and picks the tool that fits what you ask.
- You add an MCP server in your assistant with an address, for example
https://app.half5.com/mcp. - The assistant opens that program's sign-in screen. You sign in and give consent (OAuth). Your password never reaches the assistant.
- You ask something in plain language, like ‘log two hours for the login fix’.
- The assistant calls the matching tool and shows you the result.
Tools, not full access
An MCP server doesn't give the assistant free access to a database. It only offers the tools its maker built, with the rights of the account you signed in with. If your account can't see colleagues' hours, neither can the assistant.
You revoke access like you would for any connected app: in the program itself, or by removing the server from your assistant.
- Local or remote. An MCP server can run on your own computer or in the cloud. A cloud server needs no install: you paste an address and sign in.
- Open standard. Anthropic published MCP as an open specification in November 2024; OpenAI, Cursor and Microsoft support it too.
- Consent per connection. With a remote server you sign in through OAuth, the same technique as ‘Sign in with Google’.
Where it helps
MCP is most useful for work you already do in an assistant. A developer fixing a bug with Claude Code can ask the same assistant to log the time at the end. A consultant writing a report in ChatGPT can ask what's already been logged this week.
It's less suited to tasks where you want to check every field before anything is saved. So pick software that keeps everything an agent does as pending or up for approval. How that works for hours is in Logging hours with Claude, ChatGPT or Cursor.
How Half5 does it
Half5 runs its own MCP server at https://app.half5.com/mcp. You add that address in your assistant, sign in with your Half5 account and give consent. There's nothing to install.
The server offers five tools: log_work, search_customers_projects, list_entries, update_entry and delete_entry. A connected agent only sees your own hours; invoices, colleagues and settings stay out of reach. Entries from an agent arrive as ‘Not approved yet’ and are otherwise regular registrations: they show up in your overview and go through approval and invoicing like your other hours. How log_work picks the right client is in How log_work works.
Frequently asked questions
Is MCP the same as an API?
Almost. An MCP server often uses a regular API behind the scenes, but it describes its tools so a language model can choose and call them on its own. You don't have to write code to let an assistant use it.
Which AI assistants support MCP?
Among others Claude (Claude.ai, Claude Desktop and Claude Code), ChatGPT through connectors, Cursor, OpenAI Codex and VS Code. Where exactly you add a server differs per app.
Can an assistant change things through MCP on its own?
Only through the tools the server offers and only with your account's rights. Good software keeps an agent's changes up for review, so you or your manager still approve them.
Do I need to install anything?
Not for a remote MCP server: you add an address and sign in. Only local MCP servers, which run on your own computer, need installing.
