You end this page with a way to decide between an MCP server and a single API call for one specific job — not a rule that decides for you.
This is a decision, not a method, so there is no assignment here and nothing to build. It sits under Specialist 4, Automate recurring work, because skills, hooks and MCP all answer the same question — how does your method run without you — from different angles.
MCP vs API: the two axes
Two things decide this, not one. Weigh both before reaching for either.
Token cost
Every tool an MCP server offers has to describe itself — its name, what it does, what arguments it takes — and that description sits in the context window on every single turn, whether the tool gets used or not. Connect five MCP servers with a dozen tools each, and that's dozens of descriptions loaded before you've typed a word. A single API call has no such description sitting in the conversation: its shape lives in a file, not in the window.
Access scope
An MCP server hands over whatever it was built to expose — often a whole platform's worth of actions, because that's what makes it useful to many different tasks and many different people. A single API call reaches exactly the one endpoint you wrote it against, nothing more and nothing less. Wider access is convenient until it's also the thing a reviewer has to reason about.
Where the line blurs
Skills optimise the MCP side. A well-scoped skill turns a dozen tool calls someone would otherwise make by hand into one instruction — the token cost of the tool descriptions buys back the time a person would have spent chaining them manually.
Code optimises the API side. A script that makes one API call costs nothing in the context window on any given turn, because its logic lives in a file that runs, not in a description that gets read. It also does the exact same thing every time, which an MCP tool call, reasoned about fresh each turn, does not guarantee.
Which is why the line blurs in practice. A lot of real setups end up as a skill that wraps a direct API call — the ease of "one instruction, done" from the MCP side, without paying MCP's per-turn tax for every other tool sitting unused in the same server.
A worked comparison
Same job, read from each side of the choice: list a repository's open issues. Both snippets are illustrative, not code from a real build.
What an MCP tool description costs
{
"name": "list_issues",
"description": "List issues in a repository. Supports filtering by state, labels, assignee, and milestone. Returns issue number, title, state, labels, assignee, and body.",
"input_schema": {
"type": "object",
"properties": {
"owner": { "type": "string" },
"repo": { "type": "string" },
"state": { "type": "string", "enum": ["open", "closed", "all"] },
"labels": { "type": "array", "items": { "type": "string" } },
"assignee": { "type": "string" },
"milestone": { "type": "string" }
},
"required": ["owner", "repo"]
}
}
// One of dozens of tool descriptions the server exposes.
// This one loads into every turn's context whether it's called or not —
// so do the other create_issue, update_issue, add_comment, merge_pr … tools.What a single API call costs
// list-open-issues.js — the one job this script does, and nothing else.
async function listOpenIssues(owner, repo) {
const res = await fetch(
`https://api.github.com/repos/${owner}/${repo}/issues?state=open`,
{ headers: { Authorization: `Bearer ${process.env.GITHUB_TOKEN}` } },
);
return res.json();
}
// This description exists here, in the file. It costs zero tokens in the
// context window on any turn — Claude runs the script, it doesn't reason
// about a schema for it first.Laid out side by side, the trade stays consistent across every job you'd point either one at.
| MCP server | Single API call | |
|---|---|---|
| What loads every turn | Every tool the server exposes — list_issues plus create_issue, update_issue, add_comment, merge_pr and whatever else ships with it | Nothing. The request shape lives in a file that runs, not in the conversation |
| What it can reach | Whatever the server was built to expose — usually most of the platform's API surface | The one endpoint the script calls, and only that |
| Who has to maintain it | Nobody, once it's connected — the server's author keeps it current | You, or whoever wrote the script and its auth handling |
| What it costs to add a second job | Nothing new — the server likely already exposes the next tool you need | A second script, or a second function in the same one |
When each choice is right
MCP tends to fit when…
the task changes shape from one session to the next, several people need the same connection without each writing their own client, or the platform's API is large enough that hand-rolling a wrapper for every action isn't worth it. The per-turn token cost buys flexibility you'd otherwise pay for in setup time.
A single API call tends to fit when…
it's the same job, run the same way, over and over — the kind of thing a skill or a script should just do without re-reasoning about a schema each time. You're already comfortable writing and maintaining the wrapper, and you want that job to cost nothing in context no matter how many times it runs.
Neither is the default. Check your own setup with /context: if a connected server's tool descriptions show up as a visible slice of the window, that slice is what this page means by token cost — weigh it against how often you actually reach for that server's tools.
Skills are the thing that usually settles it in practice — see Specialist 4 for where skills live and how they get wired to a hook, and Claude Code skills for designers for a working set of skills that wrap direct calls instead of a broad connection.
How eliminating SaaS costs changed the MCP-versus-skills calculus
FAQ
The course
Core modules first. Specialists are optional — take the one that matches your work.
- Core 1
Stop Claude guessing
Turn a vague idea into a brief Claude can build from: facts, assumptions marked as assumptions, agreed success criteria.
- Core 2
Stop starting over
Set up project memory so a fresh session, or a teammate, can pick the work up from the files alone.
- Core 3
Move from idea to evidence
The route from brief to reviewed change: research, one shared page, a prototype, a plan, a small build, a second opinion.
- Specialist 4
Automate recurring work
Encode a procedure you repeat as a skill, and fire it automatically with a hook.
- Specialist 5
Build in parallel
Run two pieces of work at once without them blocking each other, then review how they fit together.
- Specialist 6
Run research you can trace
Research where every claim traces back to a saved source, and gaps are marked instead of filled in.
- Standalone
MCP or a single API call
A judgment call on two axes: what it costs you in context window, and how much access it opens.
You are here
Behrad Mirafshar
Founder, Bonanza Design
Founder of Bonanza Design. Builds operating brains for companies in the AI knowledge crisis. Multi-week engagements, run on the client's infrastructure, owned by the client.
Connect on LinkedIn