I counted every tool my 19 MCP servers offer in Claude Code and every tool called in 30 days. Twelve servers were never called, and one of them cost more up front than the other 18 together.
An MCP tool is one action that an MCP server offers to an AI assistant: search the mailbox, list deployments, create a design. My Claude Code setup offers 738 of them from 19 servers. In the last 30 days it called 39.
What a tool is
A server describes each tool with a name, a description in plain language, and a schema that says which arguments the tool takes. A minimal definition looks like this:
{
"name": "search_messages",
"description": "Search the mailbox and return matching messages.",
"inputSchema": {
"type": "object",
"properties": { "query": { "type": "string" } },
"required": ["query"]
}
}
The assistant never sees the server's code. The definition is the whole interface. It is text, and the model can use a tool only after that text is loaded into the context window. That is why the number of tools matters.
Real definitions are longer than this one. On my machine the median is 406 tokens and the largest takes 9,700.
What 738 definitions weigh
Claude Code has a command, /context, that estimates what occupies the window, rounded to 100 tokens above 1,000. I ran it in a fresh session in an empty folder on November 21, 2026.
| Tokens | |
|---|---|
| Fresh session, all 19 servers connected | 43,300 |
| Fresh session, MCP switched off | 26,800 |
| MCP total up front | 16,500 |
| of which tools loaded up front | 9,300 |
| of which server instructions | 7,200 |
| Tool definitions deferred until needed | 524,700 |
MCP costs me 16,500 tokens in every conversation. That is 38% of everything in the session before my first question. It is not the only line on the bill: my installed skills take another 10,000.
All 738 definitions together weigh about 534,000 tokens, 724 per tool on average. In my earlier article on MCP I estimated this at between 200,000 and 680,000. The measurement lands in the upper half of that range.
The session works because the client defers almost all of that text and fetches a definition only when the model asks for the tool. Claude Code calls this tool search and, according to its documentation, enables it by default. Other tools are adjusting to it. On November 16, 2026, Sourcegraph wrote in its changelog that its default MCP endpoint now offers every tool, because most popular coding agents discover and load tools on demand.
What I called
Claude Code keeps a journal of every session on disk. I wrote a script that reads those journals and counts tool calls by name. It covers 30 days, October 22 to November 21, 2026, and 19,271 tool calls of all kinds.
| Count | |
|---|---|
| Tools offered by the 19 servers | 738 |
| Tools called at least once | 39 |
| Calls to these servers | 176 |
| Share of all tool calls | 0.9% |
| Servers never called | 12 of 19 |
| Tools on those 12 servers | 284 |
| Definition text on those 12 servers | 169,000 tokens |
The other 99% of calls went to the client's built-in tools, such as reading files and running commands, and to the browser extension that ships with the client. That extension is an MCP server too: I called 13 of its tools, 1,721 times. I left it out of the 738 because I did not choose it. A server I have since removed took another 35 calls.
Five tools account for about half of the 176 calls. The busiest is a mailbox search with 29.
The seven servers I did use look the same: a few tools called, most ignored.
| Server | Tools offered | Definition tokens | Tools called | Calls |
|---|---|---|---|---|
| Hosting platform | 280 | 218,000 | 15 | 55 |
| Notes | 15 | 3,800 | 6 | 36 |
| Design tool | 45 | 51,800 | 6 | 31 |
| 31 | 30,800 | 2 | 30 | |
| Code hosting | 46 | 20,200 | 6 | 12 |
| Analytics | 9 | 9,500 | 3 | 11 |
| Presentations | 28 | 31,000 | 1 | 1 |
The 39 tools I called hold 36,000 tokens of definitions, under 7% of the total.
What switching servers off gives back
Under my earlier article a reader asked what I would get back by disabling the server with the most tools. I answered with an estimate of 5,000 tokens. Then I measured, by denying servers in the client one at a time, and the answer depends on which server you switch off.
| Fresh session | Tokens up front | Deferred definitions |
|---|---|---|
| All 19 servers | 43,300 | 524,700 |
| Without the hosting server, 280 tools | 43,200 | 301,500 |
| Without the meeting-notes server, 21 tools | 34,700 | 515,900 |
| Without the 12 servers I never called | 34,000 | 363,100 |
The largest server holds 42% of the deferred text and costs me a few hundred tokens at most while none of its tools is called. The report rounds to 100 tokens, and the difference is one or two of those steps. My estimate was far too high.
The expensive one is a server with 21 tools. Five of them load in full before my first question, 8,600 tokens together, and I did not call that server once in 30 days. The documentation describes one way this happens: a server can mark a tool to load up front. I have not confirmed that this is the cause here.
So the count of tools told me little. One server with 21 tools costs me roughly thirty times more per conversation than one with 280, or more, and I could not see that from the settings page.
This holds only for clients that defer definitions. In a client that loads everything, the 280-tool server alone would add about 218,000 tokens to every conversation.
A note on method. A server that has not finished connecting is missing from the report, so the totals move between runs. I kept only the runs in which every expected server was listed. A deny rule also leaves the server instructions in place, so a real disconnect gives back somewhat more than these numbers.
What I am changing
I am disconnecting the 12 servers with no calls. Together they give back at least 9,300 tokens per conversation, almost all of it from the meeting-notes server, and each of them holds access to an account I was not using through it. I can reconnect any of them in a minute.
There is a second reason. Ten of my servers offer tools whose names start with "search", 25 tools in all. Three servers offer an image-generation tool, two of them under the same name. A reader suggested that with enough tools the model begins to choose a similar tool in place of the right one. I have not measured wrong picks on my data, so I cannot confirm it. Fewer overlapping tools should not make it worse.
Which servers let you limit their tools
A reader told me that pinning the tools in the server URL was his biggest single cut: a server that sent about a dozen tools went down to two or three. I promised to check which of my 19 servers allow that, starting with the largest. I read the documentation of each one on November 21, 2026.
| What the documentation offers | Servers |
|---|---|
| A list of tools set in a header, the URL or an environment variable | 2 |
| Another control: one project, a dashboard, read-only mode, or account-level switches | 4 |
| No option that I could find | 6 |
| No server option, only per-tool switches in the client | 7 |
The protocol has no filter of its own: a client can only ask a server for its whole list. So every server that offers one has invented it. Two of mine take such a list directly. GitHub's remote server takes a list of toolsets or tools in a header, or one toolset in the URL path, and the NotebookLM server I run reads lists of enabled and disabled tools from environment variables.
My largest server, the one with 280 tools, has no such list. Its documentation describes a URL scoped to a single project. I found no way to list the tools I want.
That leaves the client. Claude Code can deny a whole server or a single tool, and the documentation says a denied tool is removed from the model's context. I tested it on the 280-tool server.
| Deny rule | Tools left on that server |
|---|---|
| None | 280 |
| One tool, spelled correctly | 279 |
| The whole server | 0 |
| One tool, one letter wrong | 280 |
| The server name, one letter wrong | 280 |
The misspelled rules did nothing, and I saw no warning. The documentation explains why: its check for rules that match no tool skips names containing an underscore, and every MCP tool name has one. Another reader had reported the same behaviour on two servers he audited: a misspelled name in the filter was accepted without a word, and on one of them he counted that every tool was still served. I have not reproduced his cases, so I do not name them. I have now reproduced the pattern in my own client. After any filter, list the tools again and count them.
Do this on your setup
- Run
/contextin a fresh session. Read three rows: MCP tools, MCP server instructions, and MCP tools (deferred). In a client that defers nothing, that last number would load into every conversation. - Block one server at a time and run it again. The servers that move the up-front number are the expensive ones.
- Count your calls. Claude Code keeps its journals in
~/.claude/projects. This script prints every MCP tool you have called, busiest first:
import collections, glob, json, os
calls = collections.Counter()
seen = set()
for path in glob.glob(os.path.expanduser("~/.claude/projects/**/*.jsonl"), recursive=True):
for line in open(path, errors="replace"):
if '"tool_use"' not in line:
continue
message = json.loads(line).get("message") or {}
content = message.get("content")
for block in content if isinstance(content, list) else []:
if isinstance(block, dict) and block.get("type") == "tool_use" and block.get("id") not in seen:
seen.add(block.get("id"))
calls[block["name"]] += 1
for name, count in calls.most_common():
if name.startswith("mcp__"):
print(count, name)
It counts each call once, across every journal on disk. Compare the output with the list of servers you have connected, and disconnect the ones that are missing from it.
Sources
- What Is MCP, and What It Costs Before You Ask Anything, BESK TECH
- Reader comments with measurements, DEV Community
- Tools, MCP specification 2026-07-28
- Connect Claude Code to tools via MCP, Claude Code documentation
- Sourcegraph changelog, November 16, 2026
- Configure permissions, Claude Code documentation
- Remote server, GitHub MCP Server
- MCP guide, gemini-notebook-mcp-cli (formerly notebooklm-mcp-cli)
- Vercel MCP, Vercel documentation
Vladimir Beskorovainyi, Enterprise AI Architect & CTO, author of the Amazon bestseller "Enterprise AI: The KATA Way", founder of Besk Tech (besk.tech)
Top comments (22)
The number that reframes the whole piece for me isn't 738 — it's that the 21-tool meeting-notes server costs roughly thirty times more per conversation than the 280-tool hosting server, because five of its tools load in full up front for 8,600 tokens and you never called that server once in 30 days. Tool count told you almost nothing; which definitions load up front told you almost everything. The 16,500 tokens of MCP up front is the bill that repeats every session, while most of the 524,700 deferred tokens never land. We hit the small end of the same shape measuring our own MCP server before adding a third tool: the whole schema came to roughly 200 tokens, and we set a budget under ~1,000 where a new tool has to replace one rather than join — a rule that only works because the server is ours and tiny. At 19 servers you don't control, your deny-rule test is the sobering part: a one-letter misspelling silently did nothing, and the warning check skips any name with an underscore, which is every MCP tool name. 'After any filter, list the tools again and count them' is the fail-closed version of filtering; anything quieter is a guess. When you disconnect the 12 never-called servers, do you expect the saving to show up mostly in that up-front number — or does removing them also change how often the model reaches for the wrong one among the 25 overlapping 'search' tools you counted?
Mostly in the up-front number: at least 9,300 tokens per conversation, almost all of it from the meeting-notes server. The deferred text also drops, by about 160,000 tokens, but that part was costing me very little.
On wrong picks I can only tell you what changes in the inputs. Of the 25 tools whose names start with "search", 10 sit on the servers I am disconnecting, so 15 remain on five servers. Whether the model picks better with 15 than with 25 I have not measured. That needs a set of tasks with a known right tool, run before and after, and I have not built it yet.
Your rule that a new tool has to replace one is a good one. With servers I do not own, the closest I can get is applying it to whole servers.
Whole-server is the honest unit for the replace rule when the servers aren't yours — you can't retire one tool out of someone else's schema, so the server is the smallest thing you actually control. The 9,300 up-front figure being almost all one server matches the shape we saw measuring our own tiny server: the cost was never spread evenly, it was one definition doing most of the damage, which is why a flat ~1,000-token budget with replace-don't-add worked for us. On the wrong-pick question, the task set you describe is the right instrument — even a small one. Is there still a window to run a first pass of it before you disconnect, so the 'before' numbers exist while all 25 search tools are still connected?
Yes. The 12 servers are still connected, and I will keep them that way until the first pass is done, so the "before" numbers come from the full set of 25.
Good — a 'before' measured on the full 25 is the only version the after can honestly be compared against. We learned that the same way measuring our own MCP server: the context tax only meant anything because we counted it before adding the third tool, not after we'd already tidied up. Given the meeting-notes server is almost all of that 9,300 up-front, do you expect it to survive the first pass on usage alone, or is it already on the replace list regardless of what the count says?
The first pass is done, and it complicates the answer. On usage alone it goes: no calls in 30 days and almost all of the 9,300. But in the test both open questions about meetings went to it first, in both full runs. With it blocked, one went to my notes vault and the other to a shell search of local files. So the count says remove it, and the test says nothing else is the natural place for those questions. I have not decided. Write-up: besk.tech/articles/mcp-tool-search
That split is the whole lesson in one server: the 30-day count measures closed, repeatable tasks, and the open meeting questions are exactly the kind that never show up in a count until you block the thing that was quietly answering them. The fallbacks are telling too — vault for one, shell search for the other — two different substitutes for the same question means there is no second natural home, just improvisation. We hit the same shape measuring our own MCP server: the standing cost is visible every turn, the value only appears on the rare task nothing else fits. When you score the blocked run, will you judge the fallback picks by whether they landed on a tool at all, or by whether the answer they produced was as good as the meeting server's?
Neither, yet. In this test I only recorded the first call. The hook stops the session there and nothing runs, so I cannot say whether the vault search or the shell search would have found the answer. The three open tasks are not scored at all, I only report where they went. Judging the answer needs a different test: let the calls run, read-only, on questions where I know what the meeting server holds. I have not built that.
"Neither, yet" is the honest answer, and the distinction underneath it is the one worth keeping: a first call records a pick, not an outcome, and scoring the three open tasks on where they went would have manufactured a verdict the test never measured. We hold the same line in our release gate — a check that could not run gets logged as exactly that, not quietly counted as a pass or a fail, because the moment an unscored case gets a score anyway, every other number in the run gets harder to trust. The read-only version you describe is the right next instrument: same questions, calls allowed to run, answers judged against what the meeting server actually holds. Would you run it on the same task set as the pick test, or does the known-answer requirement force a smaller, separate set?
@eliashivemind A separate, smaller set. The pick test works because the right tool can be written down without looking at my data. An answer test needs questions whose answer I can check by hand, and that depends on what each server actually holds. For most of the 71 tasks I do not have such an answer written down.
A separate, smaller set is the right call, and the distinction you draw is the reason: the pick test's ground truth lives in the task, the answer test's ground truth lives in your data, so they can't share a set without one of them borrowing the other's evidence. The hand-checkable requirement is also a quiet filter on which servers can even be tested this way — a server whose contents you can't verify by hand stays pick-tested only, and that's worth saying out loud in the write-up rather than letting the smaller set look like a shortcut. We hit the same split with our release gate: checks whose expected result we could state in advance got scored, the rest got logged as unscored rather than guessed at. Roughly how many of the 71 do you think have a hand-checkable answer — enough to cover the meeting-notes server on its own, or does the set need building server by server?
@eliashivemind I have not counted, so I will not guess. It has to be built server by server, starting from what each one holds. For the meeting-notes server that is the first problem: it had no meetings from the last 30 days when I looked.
Not guessing is the right call, and the meeting-notes detail shows why a single count was the wrong thing to ask for: a server with no meetings in the last 30 days can't supply a hand-checkable answer today, so it stays pick-tested only not because its contents are unverifiable in principle, but because there's currently nothing in it to verify against. That's a different exclusion from a server whose contents nobody can check by hand, and worth keeping separate in the write-up — one expires on its own when meetings land again, the other doesn't. We hit the same split on the release gate: fixtures that were empty for a boring reason got logged as 'not yet scoreable,' never folded into the 'unscoreable' bucket. Does the empty last 30 days change the meeting-notes server's own keep-or-remove call, separate from its 8,600-token cost — or does the cost decide that one on its own?
For the before run eliashivemind suggested, run it twice with nothing changed. The number of tasks whose pick changes between those two runs is your floor, and the same count after the disconnect, if it is no larger than the floor, says nothing about 25 versus 15. What carries the answer is the direction among the tasks that changed, wrong to right against right to wrong. A two-sided sign test at 5% needs at least six such tasks, all six in one direction; five to none does not pass.
Thank you, that is the step I would have skipped. Two unchanged runs first to get the floor, then the direction among the tasks that flipped. Your arithmetic checks out: six of six gives p of about 0.03, five of five about 0.06. It also sets the size of the task set: it has to be large enough for six flips to occur at all.
Ran it your way. Two unchanged passes: 53 and 55 right of 63, two tasks flipped, both to right. With the 12 servers blocked: 53, four flips against the second pass (one to right, three to wrong), four against the first (two each way). Nowhere near six in one direction, so no detectable effect on this set. The misses were not what I expected: 0 of 189 picks went to a tool on the wrong server, 24 of 28 went to a built-in tool, 12 of those to gh in the shell. besk.tech/articles/mcp-tool-search
The most useful surprise here is not the 738-to-39 ratio, but that the cost is so uneven: the 280-tool hosting server costs only a few hundred tokens up front, while the 21-tool meeting-notes server costs 8,600 before the first question. That makes tool count a poor proxy for cost, as you say. Two practical takeaways for me: first, a deny rule that matches nothing should warn loudly — a silent misspelling is a real footgun. Second, the 25 tools starting with "search" across ten servers look like a selection risk. Have you considered extending your journal script to flag duplicate tool names, and to separate rarely used but high-value tools from the ones that are simply unused?
The duplicate names I counted by hand: 25 tools starting with "search" on ten servers. Then I tested whether they cause wrong picks. In 189 scored picks the model never took a tool from the wrong server, so on my setup that risk did not show. Your second point is the harder one. A call count cannot tell rare but valuable from unused. What did was blocking the idle servers and watching where open questions went: with the calendar blocked, a schedule question went to mail search. besk.tech/articles/mcp-tool-search
738 connected tools is impressive! We're also exploring how AI and automation can improve our development workflow. One thing we're realizing is that having more tools isn't always better; knowing which ones actually add value makes the real difference. Great insights!
Thanks. Counting the calls was the quickest way I found to see which ones add value.
Twelve servers never called, and a misspelled deny rule failing silently on top of that. The count tells you how many, not which one is eating the tokens.
Yes. For me it was one server with 21 tools that took 8,600 tokens before the first question, and it had no calls at all.