Apify MCP 工具 token 開銷:10 個已啟用 Actor 的定義與結果測量
Apify MCP tool cost: 10 live Actor definitions plus the tokens in their results
作者測量 Apify 託管 MCP 伺服器上的 10 個 Actor,工具定義共 27,797 tokens,單次三列結果最高達 7,434 tokens。
I enabled ten of our Apify Actors on the hosted MCP server and counted 27,797 reference tokens in their tool definitions. Then I pulled the first three rows from a real run of each Actor. The biggest response, from our rental scraper, took 7,434 tokens. Asking for the same rows with just the fields I needed cut that to 298.
For a Claude or Cursor workflow, those are two separate levers. The tool list determines what definitions are available to the client. The dataset request determines the size of the result it can pass to the model.
Ten live tools, three rows each
I ran each Actor and used get-dataset-items on its default dataset with limit=3. Each call returned three rows at offset zero. The tool definitions came from one tools/list response with the ten named Actors enabled on https://mcp.apify.com. I used the current server and runs on November 18, 2026 UTC.
Actor under automatic/
|
Run input in brief | Tokens per tool definition | Tokens per three-row MCP result |
|---|---|---|---|
companies-house-scraper |
Search for Tesco companies | 2,049 | 1,259 |
crypto-fear-greed |
Three days with market data | 2,063 | 1,656 |
crypto-funding-rates |
BTC, ETH and SOL on Hyperliquid | 2,080 | 1,231 |
wikipedia-pageview-spikes |
English Wikipedia, November 15 | 2,434 | 2,081 |
youtube-search-scraper |
Search for SQLite tutorials | 2,533 | 1,395 |
google-maps-places-scraper |
Coffee shops in London, Ontario | 2,724 | 3,664 |
google-news-scraper |
Search for SQLite news | 2,775 | 2,510 |
trustpilot-reviews-scraper |
Chewy reviews | 3,279 | 3,348 |
treasury-yield-curve |
Daily nominal and real yields, November 9–13 | 3,526 | 6,483 |
inmuebles24-mx-scraper |
Apartment rentals in Polanco | 4,341 | 7,434 |
The 27,797 figure counts the ten definitions together as a JSON array. The complete array, including the server's four shared run and storage tools, was 31,598 tokens. The Treasury response was larger than the reviews response in this sample. Its rows include rate and change fields across maturities, so a small row limit can still produce a large result.
Same dataset, fewer fields
On the Polanco run, I kept the dataset, offset, and limit the same, and requested fields="price,currency,totalAreaM2,bedrooms". The full MCP result went from 7,434 to 298 tokens, a 96% reduction. The first listing still had null for total area and bedrooms. Projecting fields keeps the source values, including missing ones. If you're calculating price per m², ask for the area and currency as well as the price. Inmuebles24 Actor
| Field filter | Rows | Full MCP result tokens |
|---|---|---|
| All fields | 3 | 7,434 |
price,currency,totalAreaM2,bedrooms |
3 | 298 |
Apply this in Claude or Cursor
Use a workflow-specific Apify connection when you know the Actor you need. For example, https://mcp.apify.com?tools=automatic/inmuebles24-mx-scraper selects the rental Actor alongside shared run and storage tools. A smaller tool list reduces the definition payload available to a client that loads tools eagerly. For an exploratory session, broader discovery can still be useful.
After the run, ask for a preview with the fields the task needs. For example, “Read three items from the run's default dataset with the fields price, currency, totalAreaM2, and bedrooms. Keep nulls. Show the unit and source currency.” If you need a calculation over the whole dataset, run it in code and pass the summary to the model.
These are payload measurements with a common tokenizer. Your model's tokenization, tool loading, caching and handling of MCP results determine how much of that payload reaches the model and what it costs. The full result here includes both MCP text content and structured content, which carry overlapping data. Read the provider's actual usage for a billable token count.
Count it yourself
Capture the JSON-RPC response to tools/list and one response to tools/call for get-dataset-items. For an SSE response, use the JSON from the matching data: line. Save those JSON-RPC messages as tool-list.json and dataset-call.json. The response body to count is the inner result object, with the JSON-RPC envelope removed. The list and each definition are counted separately.
import json
from pathlib import Path
import tiktoken
encoding = tiktoken.get_encoding("cl100k_base")
def count(value):
wire = json.dumps(value, ensure_ascii=False, sort_keys=True, separators=(",", ":"))
return len(encoding.encode_ordinary(wire))
tools = json.loads(Path("tool-list.json").read_text())["result"]["tools"]
actors = [tool for tool in tools if tool["name"].startswith("automatic--")]
print("Actor definitions:", count(actors), "All definitions:", count(tools))
for tool in actors:
print(tool["name"], count(tool))
result = json.loads(Path("dataset-call.json").read_text())["result"]
print("Dataset rows:", len(result["structuredContent"]["items"]))
print("MCP result tokens:", count(result))
Apify run IDs
| Actor | Run ID | Dataset ID |
|---|---|---|
| Companies House | qMSAU7AHq8TBKv3QV |
D5J9OQ69VVQ9LAuk9 |
| Crypto Fear & Greed | JflT8tJhN2Q27uRB9 |
I4iUWE2Og5DJEQHfP |
| Crypto Funding Rates | jdBx5FQjy2QYqYtqb |
9VlUYZo17bxPcfN70 |
| Wikipedia Pageview Spikes | uay8F3LXeRbIdAYmD |
D24mOqRsFdiYVpZgI |
| YouTube Search | mJGdglij0uj3vQr0N |
ld4lCDerhQxhT3Z9T |
| Google Maps Places | b6tgsrLFToi5jMvGB |
sX7lng7ZWrvzrV70K |
| Google News | koGRDmJ8v34DOb8zu |
UbXzexhWkTd5oXe7v |
| Trustpilot Reviews | CUm7Cgt5c38UoGLso |
ZM9y1SWTCgLIHjrHu |
| Treasury Yield Curve | iTfAQZHtF5aCTa2Bb |
OYhE65XbI2uFZgMkl |
| Inmuebles24 Mexico | 1CzUe5PdPjy1dGdZM |
xKIJfQc3hMhxamIv3 |
Affiliate disclosure: The Apify Store link carries our referral code. If you sign up for a paid plan through it, ssh.autos may earn a commission at no cost to you.
Originally published at ssh.autos.
來源:DEV Community · dev.to