Adding an MCP server costs you a config entry, a process, some of your client's tool budget, and a new place for something to go wrong. Sometimes that is obviously worth it. Often the task is prompt-shaped and a server makes the answer worse. Four tests decide it.
A conversation forgets. If the task is "what did I bill Acme in August", something has to have been recording since August, and no prompt can produce that after the fact. Timers, invoices, expenses, deposits and ledgers are all this shape. Anything where the value is in the accumulated record needs a server.
If the task is "rewrite this paragraph", nothing needs to be remembered and a server adds nothing.
A model doing arithmetic in prose is usually right and occasionally not, and you cannot tell which from looking at the output. Money is the case where that is unacceptable. On this collection, amounts are integer minor units, each line is rounded once and then summed, and the total of a report equals the sum of the lines you can see. That is a property of the code, not of the model's attention that day.
Ask: if this number were wrong by a cent, would anyone care? If yes, it belongs in code.
Reading an xlsx, merging PDFs, writing a real .docx, making a zip, generating a QR code. A model can describe all of these and produce none of them. This is the clearest case for a server, and it is also where a hosted endpoint fails, since a remote server cannot see your disk.
A one-off conversion is faster done by hand than by installing anything. The threshold is roughly weekly. Below that, the install is the expensive part.
| Task | Verdict |
|---|---|
| Draft a client email | Prompt |
| Summarise a document you pasted | Prompt |
| Work out what you billed last month | Server: it has to have been recording |
| Make an invoice PDF | Server: a model cannot produce a file |
| Decide what to charge a new client | Prompt, informed by a server's report |
| Read a 4,000-row spreadsheet | Server: pasting it costs more than it is worth |
| Convert 80 EUR to PLN | Server, if the rate has to be defensible; prompt if it does not |
| Rename some files | Neither. Use the shell. |
Tool budgets are real. Windsurf's Cascade reaches at most 100 tools at once, and every enabled
server spends from it. The office-suite bundle here exposes every child at once, and one
tools/list against a built v0.20.0 bundle on 2026-09-07 returned 292 tools from
31 child servers. That is convenient when you do not know in advance what you will need, and
it is over every client ceiling named on this site.
A model also chooses worse from a longer list. Two servers you use daily beat twenty you installed because they were free. Count what you have before you add another:
claude mcp listEvery entry there is spending from your client's tool budget in every conversation, including the ones where you needed none of them.
Write down the question you actually have. If answering it requires a fact from before this conversation, a file on your disk, or a number that has to be exact, install something. Otherwise, ask the question.
Count tools rather than servers. The ones here run from 6 tools on per-diem and asset-register to 15 on kanban, counted in data/tools.json. Three or four servers is comfortable in every client listed on the setup pages. The 292-tool office-suite bundle is not.
In most clients, yes. Cline has a disabled flag per server, VS Code lists servers so you can stop one, and Claude Code's local scope is per directory, so a server added in one project is absent in another. That last one is a feature once you know about it and a mystery until you do.
No. It makes the model able to do things, and it makes certain answers exact rather than plausible. Those are different from being smarter and it is worth being clear about which one you needed.
One server that answers a question you have weekly. For most freelancers that is the time tracker, because the record only exists if something was recording, and the free tier gives unlimited timers and entries with reads over the last 7 days.
All MCP servers and prices · All guides · Buy the bundle $39