Best MCP Servers for Claude — and What Each One Costs You in Tokens

Every guide to MCP servers is a list of what to install. This one is about what happens to your bill afterwards, because the answer surprised us: five servers can cost more per month than the model does.

The Model Context Protocol is Anthropic's open standard for connecting Claude to the things you actually work with — your repository, your database, your design files, a browser. You install a server, Claude gains its tools, and the setup takes two minutes. What nobody mentions in the setup guide is that a tool is not free to have. Its definition — name, description, parameter schema — is sent to the model on every single request, whether you use it or not.

The number Anthropic publishes

You do not have to take a blog's word for this. Anthropic's own engineering write-up on advanced tool use states it plainly:

• A five-server setup with 58 tools consumes roughly 55,000 tokens before the conversation even starts.
• A library of 50+ MCP tools comes to about 72,000 tokens.
• Internally, before optimisation, Anthropic's own tool definitions ran to 134,000 tokens.

On a 200,000-token context window, the middle figure is more than a third of the window gone before you type a word. That part is widely discussed. The money is not.

What that costs, in dollars

One developer, 200 requests a day, Claude Sonnet 5 at published rates, a 400-token question and a 700-token answer. The only thing changing between rows is how many tool definitions ride along:

The tools cost fifteen times what the work costs. $47 of that bill is the actual questions and answers; the rest is a catalogue, re-sent 200 times a day, of capabilities that mostly sit unused.

Scale it to a team of ten on Opus 5 and the same five servers come to $17,915 a month. That is not a rounding error on an engineering budget — it is a hire.

Three fixes, in the order they pay off

1. Cache the definitions (biggest win, smallest change)

Tool definitions are the most cacheable thing in your request: byte-identical every time. Providers bill a cache read at roughly a tenth of the base input rate, and with 55k of definitions against a 400-token question, 99% of your input is cacheable.

84% off, with no change to which servers you run. Same tools, same answers, same model.

2. Load tools on demand

Anthropic's Tool Search Tool keeps the library available but sends only the schemas Claude asks for — the search tool itself costs about 500 tokens against ~72,000 for the full set, which Anthropic reports as an 85% reduction in token usage. In our arithmetic that is $923 → $179 a month for the same 50+ tools.

It also made the model better, not worse: Anthropic reports accuracy on their internal evaluation rising from 49% to 74% on Opus 4, and 79.5% to 88.1% on Opus 4.5. A shorter tool list is easier to choose from — for a model as much as for a person.

3. Let the model call tools in code

With programmatic tool calling, Claude writes code that calls the tools instead of round-tripping every call and result through the conversation. Anthropic's measured figure: average usage down from 43,588 to 27,297 tokens, a 37% reduction, with accuracy up on their benchmarks.

The servers people actually install

Ordered by how often they show up in setups, not by how much they cost — because the cost depends on how many tools each version exposes, and that changes between releases.

1. Filesystem — read, write and organise local files inside a boundary you configure. The foundational one; small tool surface.
2. GitHub — repositories, issues, pull requests, code search. The official server exposes dozens of tools, and it is the usual reason a context window fills up before work starts. Enable the toolsets you use.
3. PostgreSQL — query databases, inspect schemas, help write migrations, with permissions you set.
4. Playwright — drive a real browser: test flows, reproduce bugs, scrape a page Claude cannot otherwise see.
5. Context7 — pulls current, version-specific library documentation into the prompt, which is the cheapest fix for a model confidently using an API that changed last year.
6. Figma — reads design files so implementation can be checked against the design rather than against a screenshot.
7. Brave Search — quick live lookups when you need the web but not a full research agent.

The rule that follows from the numbers above: install the ones you use weekly, not the ones that look useful. Every server you keep "just in case" is billed on every request you make for any reason at all.

Agents make it worse, and the reason is arithmetic

In a loop, the definitions travel with every step — and each step also carries the conversation so far. Eight steps, five servers, 20 tasks a day:

Same agent, same eight steps, 72% cheaper — because in a loop you are not paying for those 55,000 tokens once, you are paying for them eight times per task. If you are sizing an agent, price the loop rather than the request: the AI Agent Cost Calculator models the growing context, the retries and the caching per call.

Work out your own number

• Count what your setup actually loads. Claude Code prints it: /context shows the token breakdown, and claude mcp list shows which servers are connected.
• Put that number in as the system prompt in the AI Cost Calculator, with your real question and answer lengths, and you have your monthly figure.
• Try it with the cached share at 90-99% to see what caching alone is worth before you disconnect anything.
• If your team is deciding between a Claude subscription and API keys, tool overhead pushes the API side up sharply — the subscription vs API calculator finds the crossing point at your usage.

Where these numbers come from

The token counts are Anthropic's, from the engineering posts linked above — 55k for five servers, 72k for a 50+ tool library, 134k internally before optimisation, ~500 for the Tool Search Tool, and the 85% and 37% reductions. Nothing here re-estimates them.

The dollar figures are ours, computed by importing the same functions that run the calculators on this site, against the providers' published rates (Anthropic, verified on the date shown on each calculator page). The assumptions:

• 200 requests a day per developer, 400 tokens of question, 700 tokens of answer.
• Tool definitions counted as system-prompt input, re-sent on every request unless cached.
• The team row is ten people at that same rate on Opus 5.
• The agent rows use eight steps, 350 tokens written per step, 600 tokens returned per tool call, a context that grows with each step, and no retries.

Where this biases the result: no retries, no failed tool calls and no second attempt at a bad plan, so the agent figures are floors. And tool definitions differ per server version — treat the categories as the unit, not any single server's exact count.

Related reading

This is one line item in a bigger picture: what AI actually costs once you run it walks through the other three — the model you pick, the length of the answers, and the loop. If Claude arrives on your card as a monthly plan instead of an API bill, it belongs in a subscription audit with everything else that renews quietly.