Pick a server by job
Too many MCP tools in the context window: scraping servers
Count the tools a scraping MCP server adds, see what Claude Code's tool search and Cursor's on-demand loading change, and size a server to the job.
By Shannon Mcbride, MCPScraper team · Published
This article was drafted with AI assistance from the sources listed below and edited and published by the MCPScraper team. Disclosure: the MCPScraper API, which this article mentions, is this site's own product.
- Count: List the server's tools and count every one, including opt-in groups.
- Trim: Leave opt-in tool groups off unless the job needs them.
- Check the client: Claude Code's tool search still loads tool names and server instructions at session start.
- Fit the job: Reading known public pages needs few tools; form flows need a browser server.
A scraping MCP server adds tools to every session it joins. Each tool is a name, a description and an input schema. In a client that loads every definition upfront, that text takes context before the first page is read. The question for a reviewer: how many tools does a server add, and how many of them does the job use.
All counts below were checked on 2026-11-14 against each project's own docs. These servers ship often. Recount before relying on a number.
Tool counts, open-source scraping servers
| Server | Default tools | Extra surface | Source |
|---|---|---|---|
| Playwright MCP | 24 core tools, plus browser_tabs | Opt-in groups through --caps: config 1, network 4, storage 17, devtools 13, vision 6, pdf 1, testing 5 | README |
| Scrapling | 17 tools ("seventeen tools" in its docs) | Not documented | MCP server docs |
| Crawl4AI (self-hosted) | 7 tools: md, html, screenshot, pdf, execute_js, crawl, ask | Not documented | self-hosting docs |
Playwright MCP with every opt-in group on adds 47 tools to the 25 of its default set, 72 in all. The table is a count, not a ranking. A larger server does more: a browser automation server has more tools than a crawler for good reason.
What a client loads upfront
The cost depends on the client as much as the server.
Claude Code. Tool search is on by default. The docs say only tool names and server instructions load at session start, and tool definitions are deferred until Claude needs them. There is no fixed per-server tool cap; the docs name the context window budget as the practical limit. Some setups load every MCP tool upfront: models earlier than the ones the docs list, Microsoft Foundry deployments hosted on Azure, and ENABLE_TOOL_SEARCH=false. The value auto loads tools upfront while their definitions total under 10% of the context window, and auto:N sets another percentage. A server set to alwaysLoad: true loads upfront too, and the docs note that each upfront tool consumes context that would otherwise be available for the conversation.
Cursor. A January 6, 2026 post on Cursor's blog describes syncing MCP tool descriptions to a folder, one per server. The agent starts with a short static context that includes the tool names and looks up the details when a task needs them. The post says some MCP servers include many tools, often with long descriptions, which can significantly bloat the context window.
Other clients. A client without deferred loading pays for every definition it loads. Check the client's own docs on the day.
Deferred loading lowers the cost of a large server. It does not remove it: names and server instructions still load at session start, and a tool the model loads during a task brings its definition with it.
Description length is part of the count
Ten short tools can cost less than three long ones. Claude Code cuts text past two default limits:
- server instructions, and the description of a tool that loads upfront: 4,096 characters;
- the description of a tool loaded through tool search: 16,384 characters.
When tool search matches a query against a tool's description, it reads only the part within the shorter limit. The docs advise keeping descriptions and instructions concise, with critical details and the words Claude would search for near the start.
Review steps
- Count. List the server's tools in the client and count every one, including opt-in groups the configuration turns on.
- Trim. Leave opt-in groups off unless the job needs them. On Playwright MCP, that means no
--capsgroup beyond the ones the task uses. - Check the client. Confirm whether it defers tool definitions, and for which models. A setting that loads a server upfront, such as
alwaysLoad: truein Claude Code, puts that server's tools back into context. - Fit the job. Clicking through a form flow needs a browser-control server and its tools. Reading known public pages does not. For coding agents, the Playwright MCP README itself says: "If you are using a coding agent, you might benefit from using the CLI+SKILLS instead." It says this approach avoids "loading large tool schemas and verbose accessibility trees into the model context".
For the read-only job, a page reader is the lighter fit. MCPScraper API is this site's own page reader. The MCP server registers 2 tools. Its MCP docs list them. Two browser-control servers are set side by side on tool count and snapshots in Chrome DevTools MCP vs Playwright MCP for reading pages.
Tool definitions are one side of the budget. Tool results are the other: client output caps are covered on fetchmcp.net in MCP fetch token budget.
Whatever the server, read only public pages, and check robots.txt, the site's terms and the provider's acceptable-use policy first.
Frequently asked questions
Do MCP tool definitions use context if the model never calls them?
In a client that loads tools upfront, yes: each definition is part of what the model is sent. With Claude Code's tool search, which is on by default, only tool names and server instructions load at session start, and tool definitions are deferred until Claude needs them.
How many tools does Playwright MCP add?
On 2026-11-14 the Playwright MCP README listed 24 core automation tools plus browser_tabs for tab management. Opt-in groups such as storage (17 tools) and devtools (13 tools) are opt-in through --caps.
How does Cursor handle servers with many MCP tools?
A January 6, 2026 post on Cursor's blog describes syncing MCP tool descriptions to a folder, one per server. The agent starts with a short static context that includes the tool names and looks up full details when a task needs them.
How do I turn off tool search in Claude Code?
Set ENABLE_TOOL_SEARCH=false to load all MCP tools upfront. The value auto loads tools upfront while their definitions total under 10% of the context window, and auto:N sets a custom percentage.
Sources
- Claude Code: Connect to tools via MCP (Documentation)
- Cursor: Dynamic context discovery (Blog)
- Playwright MCP README (Documentation)
- Scrapling MCP server docs (Documentation)
- Crawl4AI self-hosting guide (Documentation)