Best MCP servers in 2026: 15 measured for context cost and risk
We ran 15 popular MCP servers, from Context7 and GitHub to Playwright, Supabase and Notion, and measured what each loads into an agent's context before it does any work. Our picks by job, the numbers behind them, and how many of their tools can change things.

The best MCP servers for most developers in 2026 are Context7 for library docs, GitHub's official server, Playwright for the browser, the Filesystem server, Supabase for a database, and Exa or Firecrawl for the web. We ran each of them, plus nine others, and measured what they cost before your agent does anything: the tool definitions they load into every conversation.
That cost varies more than most lists let on. Context7 adds 2 tools and 4,945 characters. Notion's local server adds 24 tools and 74,690 characters. Every "best MCP servers" page we read ranks servers by usefulness and maintenance, and none of them counts this, so we did.
How we tested
On November 16, 2026, we started each server on our machine with the official MCP TypeScript SDK (1.32.2) as the client, called tools/list, and measured what came back:
- Tools. How many tools the server lists with its default settings.
- Characters. The length of each tool's name, description and input schema, written out as JSON. That is roughly what a client sends the model so it knows the tools exist.
- Marked read-only. Tools the server itself marks with
readOnlyHint, a promise that the tool changes nothing. Many clients run those without asking you.
GitHub's server ran with a real token. The others list their tools without a valid key, so we gave them a placeholder. This is the core of the script:
import { Client } from "@modelcontextprotocol/sdk/client/index.js";
import { StdioClientTransport } from "@modelcontextprotocol/sdk/client/stdio.js";
const [, , command, ...args] = process.argv;
const client = new Client({ name: "measure", version: "1.0.0" });
await client.connect(new StdioClientTransport({ command, args }));
const tools = [];
let cursor;
do {
const page = await client.listTools(cursor ? { cursor } : {});
tools.push(...page.tools);
cursor = page.nextCursor;
} while (cursor);
const sent = tools.map((t) => ({ name: t.name, description: t.description ?? "", input_schema: t.inputSchema }));
console.log({
tools: tools.length,
characters: JSON.stringify(sent).length,
readOnly: tools.filter((t) => t.annotations?.readOnlyHint === true).length,
});
await client.close();Run it as node measure.mjs npx -y @upstash/context7-mcp and swap in any server's start command.
15 MCP servers, measured
Sorted by how much text each one loads. Versions are the ones we ran on November 16, 2026.
- Fetch 2026.11.11: 1 tool, 1,347 characters, 1 marked read-only
- Time 2026.11.11: 3 tools, 2,094 characters, 3 marked read-only
- Sequential Thinking 2.1.1: 1 tool, 2,231 characters, 0 marked read-only
- Exa 3.4.3: 3 tools, 4,762 characters, 3 marked read-only
- Context7 4.7.0: 2 tools, 4,945 characters, 2 marked read-only
- Git 2026.11.11: 14 tools, 6,189 characters, 8 marked read-only
- Memory 2.1.1: 9 tools, 8,057 characters, 3 marked read-only
- Filesystem 2.3.0: 16 tools, 13,806 characters, 10 marked read-only
- Playwright 0.0.85: 25 tools, 18,507 characters, 7 marked read-only
- Supabase 0.17.0: 36 tools, 29,625 characters, 20 marked read-only
- Chrome DevTools 1.13.0: 33 tools, 30,318 characters, 8 marked read-only
- Brave Search 2.1.4: 8 tools, 33,192 characters, 0 marked read-only
- GitHub 2.1.0: 46 tools, 49,454 characters, 27 marked read-only
- Notion 2.6.1: 24 tools, 74,690 characters, 12 marked read-only
- Firecrawl 3.40.0: 32 tools, 76,492 characters, 22 marked read-only
Two things stand out. Tool count is a poor guide to size: Brave Search has 8 tools and still loads more text than Supabase's 36, because each tool carries a long description and schema. And Notion's local server spends about 3,100 characters per tool, since its tools are generated from Notion's full API schema.
The best MCP servers by job
Library docs: Context7
Context7 looks up current documentation for a library and puts it in front of the model, so the agent writes code against the version you use instead of the one it remembers. Two tools and 4,945 characters make it the cheapest useful server in our test. If you add one server to a coding agent, start here.
Code and pull requests: GitHub's MCP server
GitHub's official server covers issues, pull requests, reviews, branches and files. Its defaults load 46 tools and 49,454 characters, the third largest in our table. You can shrink it with flags the server documents itself:
--read-onlyleaves 27 tools and 28,049 characters, and nothing that can change a repository.--toolsetspicks groups of tools.--toolsets allgoes the other way: 91 tools and 104,380 characters.
If your agent only reads code and opens issues, run it read-only and add write toolsets when you need them.
The browser: Playwright, or Chrome DevTools for debugging
Playwright's server drives a real browser through its accessibility tree: 25 tools, 18,507 characters. Use it for filling forms, clicking through flows and checking that a page works. Chrome DevTools' server is larger, 33 tools and 30,318 characters, and earns it when you need console messages, network requests and performance traces. Pick one. With both installed, the agent sees two overlapping sets of browser tools.
Local files and git: Filesystem and Git
The Filesystem server reads and writes inside folders you list when you start it: 16 tools, 13,806 characters. The Git server adds status, diffs, commits and branches for one repository: 14 tools, 6,189 characters. Coding agents such as Claude Code, Codex and Cursor already read files and run git themselves, so these two matter most in chat apps like Claude Desktop that can't.
A database: Supabase
Supabase's server manages projects, runs SQL, applies migrations and reads logs: 36 tools, 29,625 characters. Its --read-only flag drops it to 22 tools and 15,896 characters, all marked read-only. Start there. An agent that can apply migrations to production is one bad guess away from a long night.
The web: Exa for search, Firecrawl for scraping
Exa's server returns search results and page text with 3 tools and 4,762 characters. Firecrawl's server does far more, with scraping, crawling, page monitors and browser actions, and loads 32 tools and 76,492 characters for it, the most in our test. Install Firecrawl when the agent needs to crawl sites or read pages that block plain requests. For a quick lookup, Exa or the 1-tool Fetch server costs a fraction of that.
Notes and docs: Notion
Notion's local server reaches pages, blocks, comments and data sources through the Notion API. It is also the heaviest per tool we measured, 74,690 characters for 24 tools. Notion runs a hosted MCP server as well, which we didn't measure, so compare the two before you settle on one.
What a typical stack loads before the first message
Take a common setup for a developer: GitHub, Playwright, Context7, Filesystem, Supabase and Notion. That is 149 tools and 191,027 characters of tool definitions. At around 4 characters per token, that's close to 48,000 tokens. We counted characters, not tokens, so treat the token figure as an estimate. Our MCP vs API compares one GitHub request both ways, and what each costs an agent in tokens.
Some clients cut this down. Claude Code's tool search can hold MCP tool definitions back and load them when the agent searches for one. Check what your client does before you install a tenth server. If it loads everything up front, every server you add makes each conversation more expensive and gives the model more tools to choose from.
How many of these tools can change things?
Across the 15 servers, 126 of 253 tools are marked read-only. The other 127 can write, send or delete, or at least don't promise not to. What happens when the agent calls one is up to your client: some ask first, some have a setting that stops asking, and the server itself usually doesn't check.
The same split shows up across apps in general. Corsair labels every operation in its catalog as read, write or destructive. On November 16, 2026, the catalog had 624 apps and 23,317 operations. 3,245 of those operations, 13.9%, are destructive, and 376 of the 624 apps, 60%, have at least one. Those are Corsair's own labels, but the point holds whatever labels you use. Most apps an agent can reach can also delete something.
One server for your SaaS apps
The servers above are one per tool, and that works for local things like files, git and a browser. For apps that live online, such as Gmail, Google Drive, Slack, Linear, HubSpot or Stripe, one server per app means one config block, one token and one more set of tool definitions each.
Corsair MCP takes a different approach. You add one address, https://mcp.corsair.dev/mcp, connect apps to it, and the agent sees the same short list of tools however many apps are connected. In the version checked into Corsair's repository on November 15, 2026, the full set is 14 tools and 13,397 characters, less than the Filesystem server alone. The agent finds an app's actions with list_operations when it needs them and runs them with run_script. Our own team's connection has 25 apps on it, and its tool list doesn't grow with them.
Each connection has a mode that Corsair enforces on the server, whatever the client allows:
- Read. Reads only. Writes are refused.
- Cautious, the default. Writes go through. Destructive actions, such as a delete, wait for your approval.
- Strict. Every write waits for approval. Destructive actions are refused.
- Open. Everything, without asking.
An approval shows exactly what will run and with which arguments, and covers that one action once. Agents never see the app's OAuth tokens or API keys. To add it to Claude Code:
claude mcp add --transport http corsair https://mcp.corsair.dev/mcpThen run /mcp in Claude Code to sign in.
Where this approach falls short
A short tool list has a cost. The agent spends an extra call looking up an app's operations before it can use one, where a dedicated server would hand it the tool directly. For an app you use all day in one client, a good dedicated server can be the simpler choice. And Corsair doesn't replace local servers: it reaches online apps, not your files, your git repository or a browser on your machine. Keep Filesystem and Playwright for those.
Our test has limits too. It measures what servers load, not how well they work. We ran default settings on local servers only, so hosted servers such as Linear's, Sentry's and Notion's hosted one aren't in the table. Stripe's local server timed out without a real key, so it isn't either.
One MCP address for your apps
If you want your online apps behind one MCP address with approvals before anything destructive, claim a handle at lug.dev and add its one URL to Claude, Cursor, ChatGPT or Codex. Teams building agents into their own product can start with the Corsair MCP docs.
FAQs
What are the best MCP servers for Claude Code?
Context7 for library docs, GitHub's server for pull requests and issues, and Playwright for testing in a browser. Claude Code already reads files and runs git, so it doesn't need the Filesystem or Git servers. Add a database server such as Supabase in read-only mode if your work touches one.
How many MCP servers is too many?
There's no fixed number, but measure the total. Six common servers in our test added 149 tools and about 191,000 characters, roughly 48,000 tokens, before the first message. If your client loads every tool up front, cut servers you don't use each week, or switch on the server's own read-only or toolset options.
Are MCP servers safe to install?
A local MCP server runs with your user's permissions, so it can do anything you can in the folders and accounts you give it. Install servers from the app's own company or a maintainer you trust, give them the narrowest token that works, and use read-only modes where they exist. Half the tools in our test can change something.
Do MCP servers use tokens when the agent doesn't call them?
Usually, yes. Most clients send every tool's name, description and schema to the model at the start of each conversation, so a server costs tokens even if no tool is called. Some clients, such as Claude Code with tool search, defer those definitions and load them on demand.
What are the best free MCP servers?
Every server in our test is free to run. Some reach paid services: Exa, Firecrawl and Brave Search need an API key, and their free tiers have limits. Context7, Playwright, Chrome DevTools, Filesystem, Git, Fetch and Time need no key at all.
What's the difference between a local and a remote MCP server?
A local server runs on your computer, started by your client from a command such as npx, and any key it needs sits in your config file. A remote server runs on the internet, you add it by URL, and you sign in with OAuth. Local servers suit files, git and a browser on your machine. Remote ones suit online apps and follow your account across clients.