mrkeyoor.com_
Tue 01 Sept 17:44 UTC

Best MCP Servers (2026) — we booted every one

We started all 51 servers in our set on a clean machine, ran initialize and tools/list over stdio, and wrote down what came back; the numbers are ours. Twenty of them never reached initialize — they are listed at the bottom, because the failure is the finding. The ranking below is one measured ratio: tools exposed per 1,000 tokens of tools/list, estimated at 4 characters per token. That payload is the Context Tax: you pay it on every turn whether the model calls a tool or not. Archived servers, the reference test server, and servers exposing fewer than two tools are excluded from the ranking.

Every server started over stdio on a clean machine, then ranked by capability per context token — tools exposed per 1,000 tokens of tools/list.

01

Browserbase

Browserbase

Browserbase MCP is the Browserbase-built bridge between an AI client and a cloud Chrome session running on their infrastructure. It exposes six tools: start, navigate, act, observe, extract, end. Instead of driving a local browser, your agent rents a remote one, so page loads, captchas and proxies happen on Browserbase hardware rather than your laptop.

Measured BOOTS ✓ · boot 21.7s · 6 tools · context tax ~393 tokens · 15.3 tools / 1k tokens · 0 resources · protocol 2025-06-18

Verdict Use it if you already pay Browserbase, or intend to..

Pick it if You want browser automation your agent can call without a Chrome binary, X server or headless-shell install on the machine running the client

Full review →
02

should-i-use (ours)

MrKeyoor

should-i-use is our own MCP server, published by MrKeyoor as the npm package should-i-use-mcp. It gives a coding agent six tools for library decisions: audit_dependencies, pick_library, how_do_i, docs_link, should_i_use and alternatives. It answers over stdio, needs no API key, and reports itself as version 0.2.0.

Measured BOOTS ✓ · boot 7.9s · 6 tools · context tax ~666 tokens · 9.0 tools / 1k tokens · protocol 2025-06-18

Verdict Use it if your agent keeps guessing about packages and you want a second opinion that arrives as a tool call instead of a web search..

Pick it if Your agent picks npm packages on its own and you would rather it consulted something than recalled something

Full review →
03

Git

Anthropic (reference)

Git is Anthropic's reference MCP server for working with local repositories. It exposes status, diff, log, show, add, commit, branch, create-branch, checkout and reset as callable tools over stdio, each scoped to a repository path the agent passes in. It lives in the modelcontextprotocol servers repo, ships on PyPI, and runs with uvx.

Measured BOOTS ✓ · boot 2.1s · 12 tools · context tax ~1,477 tokens · 8.1 tools / 1k tokens · protocol 2025-06-18

Verdict Use it, with one guardrail..

Pick it if Your agent host blocks raw shell but you still want it reading diffs and writing commits

Full review →
04

Time

Anthropic (reference)

Time is Anthropic's reference MCP server for clocks and timezones. It ships two tools: get_current_time, which returns the current time in an IANA timezone, and convert_time, which moves a wall-clock time from one zone to another. No account, no API key, no network calls at runtime. You run it locally with uvx as mcp-server-time.

Measured BOOTS ✓ · boot 2.0s · 2 tools · context tax ~300 tokens · 6.7 tools / 1k tokens · protocol 2025-06-18

Verdict Use it, and stop thinking about it..

Pick it if Your agent schedules things across timezones and you are tired of it inventing what time it is in Berlin

Full review →
05

Playwright MCP

Microsoft

Playwright MCP is Microsoft's browser-automation server for the Model Context Protocol. It gives a model a real Chromium session and 24 tools for driving it: navigate, click, type, fill forms, upload files, read console messages and network requests, take screenshots, and run arbitrary page code. It talks to the browser through Playwright's accessibility snapshot rather than pixels.

Measured BOOTS ✓ · boot 5.4s · 24 tools · context tax ~4,626 tokens · 5.2 tools / 1k tokens · protocol 2025-06-18

Verdict Use it, and expect to pay for it in context..

Pick it if You need the model to log in, click through a flow, and read what came back, not just GET a URL

Full review →
06

Filesystem

Anthropic (reference)

Filesystem is Anthropic's reference MCP server for giving a model scoped read and write access to local directories. You pass it a list of allowed paths on the command line, and it exposes 14 tools for reading, writing, editing, moving, listing and searching inside those paths. It refuses anything outside them, which is the whole design.

Measured BOOTS ✓ · boot 9.0s · 14 tools · context tax ~3,243 tokens · 4.3 tools / 1k tokens · protocol 2025-06-18

Verdict Use it..

Pick it if You want an agent editing files in one project directory and nothing else, enforced by the server rather than by a prompt

Full review →
07

Excel

community (negokaz)

Excel is a community MCP server from negokaz that lets an assistant read and edit real .xlsx workbooks over stdio: describe the sheets, read a range, write cells, copy a sheet, create a table, apply formatting. It runs through npx, touches files on your disk directly, and needs no account, no API key and no cloud service.

Measured BOOTS ✓ · boot 8.9s · 6 tools · context tax ~1,406 tokens · 4.3 tools / 1k tokens · protocol 2025-03-26

Verdict Use it if your spreadsheets already live as files on the machine the agent runs on..

Pick it if Your workbooks sit on the same box as the agent and you want cell-level reads without exporting to CSV first

Full review →
08

Kubernetes

community (Flux159)

The Kubernetes MCP server, mcp-server-kubernetes by Flux159, hands an AI agent your kubectl. It is a community project, not a CNCF or vendor release, and it wraps get, apply, patch, delete, logs, describe, scale, rollout, exec, port-forward and Helm install into MCP tools that run against whatever kubeconfig context the process can see.

Measured BOOTS ✓ · boot 26.3s · 23 tools · context tax ~5,964 tokens · 3.9 tools / 1k tokens · 5 resources · protocol 2025-06-18

Verdict Use it if you want an agent doing read-heavy cluster triage, and only if you have pointed it at a context where the worst possible command is survivable..

Pick it if You want an agent to read logs, describe pods and explain why a deployment is stuck, against a context scoped to one non-critical namespace

Full review →
09

Exa

Exa Labs

Exa is a search MCP server from Exa Labs that gives an agent two things: a web search backed by Exa's neural index, and a fetch tool that pulls a single URL's contents back as text. It runs over stdio from npx, identifies itself as exa-search-server, and needs an Exa API key.

Measured BOOTS ✓ · boot 13.0s · 2 tools · context tax ~534 tokens · 3.7 tools / 1k tokens · 1 resources · protocol 2025-06-18

Verdict Use it if you want your agent to search the live web and you're fine paying Exa for the index..

Pick it if You want a research agent that can find pages it wasn't told about, and you already have or will buy an Exa API key.

Full review →
10

Memory

Anthropic (reference)

Memory is Anthropic's reference MCP server for persistent recall. It keeps a knowledge graph of entities, relations and observations in a local JSON file, and exposes nine tools an assistant can call to write facts down and read them back in a later session. It ships as @modelcontextprotocol/server-memory and runs over stdio.

Measured BOOTS ✓ · boot 5.1s · 9 tools · context tax ~2,688 tokens · 3.3 tools / 1k tokens · 1 resources · protocol 2025-06-18

Verdict Use it, but understand you are buying a note-taking habit, not a memory..

Pick it if You want your assistant to remember your stack, your preferences and your colleagues' names across sessions, and you're willing to write the system prompt that tells it when to save

Full review →
11

MongoDB

MongoDB

MongoDB's own MCP server, published as the npm package mongodb-mcp-server. It gives an assistant 31 tools over a MongoDB deployment: find, aggregate, insert-many, update-many, index management, collection and database stats, plus Atlas local deployment control and a search-knowledge tool that queries MongoDB's own documentation. It speaks stdio JSON-RPC and connects with a standard connection string.

Measured BOOTS ✓ · boot 52.2s · 31 tools · context tax ~9,462 tokens · 3.3 tools / 1k tokens · 2 resources · protocol 2025-06-18

Verdict Use it if you actually work in MongoDB every day and are willing to spend the context..

Pick it if You are debugging real queries and want the model to run explain and collection-schema against the live deployment instead of guessing at your document shape.

Full review →
12

Netlify

Netlify

Netlify's own MCP server, shipped as @netlify/mcp, hands an agent nine tools for driving a Netlify account: readers and updaters for projects, deploys and extensions, readers for team and user info, and one tool that feeds Netlify coding rules back to the model. It runs over stdio and leans on whatever Netlify login already exists on the machine.

Measured BOOTS ✓ · boot 33.8s · 9 tools · context tax ~3,083 tokens · 2.9 tools / 1k tokens · protocol 2025-06-18

Verdict Use it if you actually deploy to Netlify and want the agent to check a build, flip a project setting, or install an extension without you tabbing to the dashboard..

Pick it if Netlify is your actual host and you want an agent to read deploy state and logs instead of you opening the deploys tab

Full review →
13

Firecrawl

Firecrawl

Firecrawl's MCP server, published by the Firecrawl team as the npm package firecrawl-mcp, gives an AI agent web scraping, crawling, search, structured extraction, page monitoring and a small research toolkit over the hosted Firecrawl cloud. We started it and it registered 25 tools, reporting itself as firecrawl-fastmcp version 3.24.0 on protocol 2025-06-18.

Measured BOOTS ✓ · boot 16.3s · 25 tools · context tax ~9,077 tokens · 2.8 tools / 1k tokens · protocol 2025-06-18

Verdict Use it if scraping is central to what your agent does, because the tools work and the keyless mode lets you try it before paying anyone..

Pick it if Your agent's main job is reading the live web and you want markdown back instead of raw HTML you have to clean up

Full review →
14

Tavily

Tavily

Tavily is a hosted web-search API for agents, and tavily-mcp is the vendor's stdio bridge to it. It exposes five tools: search, extract, crawl, map, and research. An agent gets ranked web results and cleaned page text back as JSON instead of raw HTML, so it never has to parse a search results page itself.

Measured BOOTS ✓ · boot 6.7s · 5 tools · context tax ~1,924 tokens · 2.6 tools / 1k tokens · protocol 2025-06-18

Verdict Use it if you want your agent to read the live web and you're willing to pay Tavily per call..

Pick it if Your agent answers questions about things that happened after its training cutoff and you want cleaned text, not HTML soup

Full review →
15

n8n

community (czlonkowski)

n8n-mcp is a community server by czlonkowski that hands your agent the n8n automation platform's node documentation. It searches nodes and workflow templates, fetches a node's full parameter reference, and validates a node config or a whole workflow JSON before you paste it into n8n. It talks to docs, not to your live n8n instance.

Measured BOOTS ✓ · boot 15.0s · 7 tools · context tax ~2,942 tokens · 2.4 tools / 1k tokens · 2 resources · protocol 2025-03-26

Verdict Use it if you write n8n workflows with an assistant and are tired of hallucinated node parameters, because validate_workflow catching a bad JSON before n8n does is worth the seven tool slots..

Pick it if You build n8n workflows regularly and want the agent to check node parameter names against the real reference instead of guessing them

Full review →
16

Desktop Commander

community

Desktop Commander is a community MCP server from wonderwhy-er that hands a model your actual machine: read and write files, apply surgical edits, run shell commands, and drive long-lived processes over stdio. We started version 0.2.47 and it advertised 26 tools plus 2 resources and no prompts.

Measured BOOTS ✓ · boot 55.1s · 26 tools · context tax ~15,226 tokens · 1.7 tools / 1k tokens · 2 resources · protocol 2025-06-18

Verdict Use it if you genuinely want an agent with a shell and a filesystem, and you accept both prices..

Pick it if You want one server that covers file edits and shell processes instead of bolting a filesystem server to a terminal server

Full review →
17

Context7

Upstash

Context7 is a documentation-lookup MCP server from Upstash. It gives a coding agent two tools: one that turns a library name into a Context7 library ID, and one that queries current docs for that library. It runs over stdio via npx, needs no credentials to start, and reported itself as Context7 version 4.0.2 when we booted it.

Measured BOOTS ✓ · boot 7.8s · 2 tools · context tax ~1,215 tokens · 1.6 tools / 1k tokens · 0 resources · protocol 2025-06-18

Verdict Use it..

Pick it if Your agent keeps generating API calls that were removed in a major version and you'd rather it look them up than guess.

Full review →
18

Pinecone

Pinecone

Pinecone's own MCP server, published as @pinecone-database/mcp, gives an assistant stdio access to a Pinecone vector database: list and describe indexes, create an index backed by a hosted embedding model, upsert records, run semantic and cascading search, rerank results, and query Pinecone's documentation. It reported itself as pinecone-mcp version 0.3.0 when we started it.

Measured BOOTS ✓ · boot 8.2s · 9 tools · context tax ~6,165 tokens · 1.5 tools / 1k tokens · protocol 2025-06-18

Verdict Use it if Pinecone is already where your vectors live and you want an agent to build and query indexes without you writing glue..

Pick it if You are building a retrieval pipeline on Pinecone and want the agent to create the index, upsert records and check describe-index-stats without leaving chat

Full review →
19

Notion

Notion

Notion MCP is the stdio server published by Notion as @notionhq/notion-mcp-server. It exposes the public Notion REST API to an agent as 24 tools: search, page read and write in markdown, block children, data source query and update, comments, and users. You bring your own integration token, and the agent can only touch pages that token has been shared with.

Measured BOOTS ✓ · boot 8.9s · 24 tools · context tax ~19,054 tokens · 1.3 tools / 1k tokens · protocol 2025-06-18

Verdict Use it, but not with every tool switched on..

Pick it if Your team's specs and meeting notes live in Notion and you want the agent reading them instead of you pasting page dumps into chat

Full review →
20

CircleCI

CircleCI

CircleCI's own MCP server, published as @circleci/mcp-server-circleci, gives an assistant stdio access to your CircleCI account: pipeline status, build failure logs, flaky test detection, artifacts, reruns, rollbacks and usage data. It boots and lists 13 tools. It also prints a deprecation banner on startup, pointing at CircleCI's hosted MCP server instead.

Measured BOOTS ✓ · boot 17.5s · 13 tools · context tax ~12,932 tokens · 1.0 tools / 1k tokens · protocol 2025-06-18

Verdict Use it only if you are pinned to a self-hosted stdio setup and you accept that the package tells you, in a box drawn with line characters, that it will stop getting security fixes..

Pick it if You debug red builds through an assistant and want it to pull failure logs and test results itself instead of you pasting them

Full review →
21

Shopify Dev

Shopify

Shopify Dev MCP is a stdio server published by Shopify as @shopify/dev-mcp. It gives a coding agent five tools for building on Shopify: one that loads API context on demand, three validators for GraphQL, Polaris components and theme files, and a documentation chunk search. It talks to Shopify's public docs, not to your store.

Measured BOOTS ✓ · boot 60.8s · 5 tools · context tax ~6,590 tokens · 0.8 tools / 1k tokens · protocol 2025-06-18

Verdict Use it if you write Shopify code and are tired of your agent inventing GraphQL fields that were removed two API versions ago..

Pick it if You are writing Shopify Admin GraphQL and want validate_graphql_codeblocks to reject a bad field before you burn a deploy finding out

Full review →

Did not boot (20 of 51)

These never reached initialize on a clean machine with no credentials supplied. Most want an API key or token before they will speak the protocol at all, which is a setup cost worth knowing before you add one to a config file.

ServerVendorWhat happened
Airtablecommunity (domdomegg)exited with code 1 before initialize completed · needs 1 env var
Apify ActorsApifyexited with code 1 before initialize completed · needs 1 env var
Azure MCPMicrosoftexited with code 1 before initialize completed
Blendercommunityexited with code 1 before initialize completed
Brave Search (archived)Anthropic (archived)exited with code 1 before initialize completed · needs 1 env var
ChromaChromaexited with code 1 before initialize completed
DuckDuckGo Searchcommunityexited with code 1 before initialize completed
ElasticsearchElasticexited with code 1 before initialize completed
Figma (Framelink)communityboot timeout after 90s (npx install + start)
GitLab (archived)Anthropic (archived)exited with code 1 before initialize completed · needs 1 env var
Google Drive (archived)Anthropic (archived)exited with code 1 before initialize completed
Google Maps (archived)Anthropic (archived)exited with code 1 before initialize completed · needs 1 env var
HerokuHerokuexited with code 1 before initialize completed
Obsidiancommunityexited with code 1 before initialize completed · needs 1 env var
PayPalPayPalexited with code 0 before initialize completed
QdrantQdrantexited with code 1 before initialize completed
SentrySentryexited with code 1 before initialize completed · needs 5 env vars
Slack (archived)Anthropic (archived)exited with code 1 before initialize completed · needs 2 env vars
StripeStripeexited with code 1 before initialize completed · needs 1 env var
SupabaseSupabaseexited with code 1 before initialize completed · needs 1 env var

See also