-
Notifications
You must be signed in to change notification settings - Fork 4.7k
mcp servers on github
Most of MCP is public code. That is genuinely useful and it also means the ecosystem is full of weekend demos that look identical to production servers from a search result. This page is where things are, and how to tell them apart in about two minutes.
The protocol itself is specified at modelcontextprotocol.io, with the
specification, the SDKs and reference servers published under the
modelcontextprotocol organisation. If you are implementing rather than
installing, start there rather than from a blog post: the SDKs exist for
several languages and the protocol methods are documented rather than inferred.
Vendor servers live with their vendor: Microsoft publishes the Playwright one, the Chrome DevTools team publishes theirs. Playwright MCP vs Chrome DevTools MCP compares those two, which are the largest browser-adjacent ones.
Community servers are everywhere else, and outnumber the first two categories by a wide margin.
Directories aggregate all of the above. They are a decent way to discover and a poor way to judge: listing is not curation, and a directory entry tells you nothing about whether the thing still works.
1. Last commit, not stars. Stars accumulate and never decay; a server with eight thousand stars and no push in a year is a worse dependency than one with two hundred pushed last week. MCP moved fast enough in 2025 and 2026 that a year-old server may not speak the current protocol revision.
2. Who owns it. A server wrapping a product, published by that product's own organisation, is maintained as a side effect of the product existing. A community wrapper is maintained while its author is interested.
3. The tool list, before you register it. tools/list is a protocol
method, so you do not have to read the source to know what a server can do -
any client will enumerate it. Look for tools that write, delete or send, and
decide deliberately rather than discovering later.
4. The size of the surface. Tool descriptions ride in the model's context on every turn, argument schemas included. Measured on our own server: 16 tools, 9,145 characters of description, 3,192 tokens per turn. A server advertising a hundred tools is asking for several times that, forever. How many MCP tools is too many has the arithmetic.
Whether it handles concurrency. Two clients, one server, one profile directory is the most common first-hour failure in the browser category and it has nothing to do with code quality: browser is already in use.
What it does when something fails. A demo returns a stack trace to the model. A production server returns something the model can act on.
Whether its claims were tested. Several servers in the stealth corner advertise capabilities we could not verify, which is stated plainly in stealth MCP servers compared along with what our own does not do.
Where is the official MCP GitHub organisation? Under
modelcontextprotocol, alongside the specification site, with SDKs and
reference servers.
How do I find MCP servers on GitHub? Directories and topic searches both work for discovery. Neither judges: apply the four checks above.
Are community MCP servers safe? They run with whatever access you grant and
expose tools that act. Read tools/list before registering, and prefer
read-only where the choice exists.
Which MCP server should I start with? Depends on the capability you want. How to choose among MCP servers maps the categories.
Is there an official list of MCP servers? The protocol organisation publishes reference servers; third-party directories cover the rest with no curation guarantee.
See also: the MCP server, and MCP tools, resources and prompts for what a server can expose in the first place.
-
The Model Context Protocol documentation, retrieved 2026-09-11, for the protocol operations including
tools/listand for the SDK and reference-server layout. - microsoft/playwright-mcp and ChromeDevTools/chrome-devtools-mcp, retrieved 2026-09-10, as the two vendor-published examples named above.
- This project's own server source, read 2026-09-13, for the tool-surface measurement.
Written while maintaining a community MCP server, which is the category this page tells you to check hardest. The four checks apply to ours: read our tool list before registering it.
- OpenAI Operator alternatives
- Open-source Operator-style agents
- Is OpenAI Operator still available?
- OpenAI Operator vs Claude computer use
- browser-use alternatives
- Choosing an AI browser agent
- Open-source AI browser agents
- Open-source computer-use agents
- What is an AI web agent?
- AI browser agents vs traditional scraping
- Cloud browser infrastructure for AI agents, explained
- Browserbase alternatives
- Firecrawl vs an AI browser agent
- Skyvern alternatives
- Stagehand vs browser-use
- Project Mariner is gone: what replaced it
- Manus alternatives
- Gemini computer use vs Claude computer use
- AIHawk, reviewed honestly by its own wiki
- AI browser vs AI browser agent: which one do you want?
- AI browser agent vs RPA: which one fits the job
- AI browser agent vs n8n, Zapier and Make
- Vercel agent-browser alternatives, compared honestly
- What is an agentic browser? Definition and the two kinds
- Open-source agentic browsers: the three layers, compared
- Choosing an MCP server for browser automation: four axes
- Stealth MCP servers compared: Camoufox, nodriver, Patchright
- Playwright MCP alternatives, and the three you don't need
- Autonomous browser agents: the four rungs of autonomy
- What is actually free in the AI browser agent stack
- browser-use on GitHub: what the repo actually gives you
- Playwright MCP vs Chrome DevTools MCP: different jobs
- How to choose among MCP servers: a map by category
- Which MCP servers are worth adding to Claude Code
- MCP on GitHub: finding servers and judging them fast
- MCP vs an API: the decision, and what the wrapper costs
- MCP alternatives: when the protocol is the wrong shape
- Why does my AI agent get blocked?
- The timing signal AI agents give off
- Agent retry loops trip rate limits, not fingerprints
- Claude computer use detected as a bot
- browser-use getting blocked: what you can and cannot change
- Playwright MCP session blocked: four causes, four fixes
- Playwright MCP and captchas: what actually gets you past
- Cloudflare and a browser MCP server: what is being read
- Can an AI agent solve a captcha? The honest answer
- Getting an AI agent to fill out forms
- Which model to use with AIHawk
- Browser problem or model problem?
- Running AIHawk's browser from Claude Code
- Extracting data to a CSV with an AI agent
- Monitoring a page for changes with an AI agent
- Running AIHawk's browser from Claude Desktop
- Running AIHawk's browser from Cursor
- Using an AI agent to hunt for apartments
- Getting website data into Google Sheets with an AI agent
- Using an AI agent to download invoices from portals
- AI agents for web research
- Using an AI agent to test your own website
- Running AIHawk's browser from Cline
- Posting to social media with an AI agent
- Posting to Facebook with an AI agent
- Posting to Instagram with an AI agent
- Posting to X with an AI agent
- Automating LinkedIn posts: read this first
- Appointment bots: what they are and what an agent can legitimately do
- Track prices across sites with an AI agent
- Build a lead list with an AI browser agent
- Run an AI browser agent on a schedule
- AI browser agent with a local LLM: what changes
- Should you log your AI agent into your accounts?
- How to write a task an AI browser agent can follow
- Move data between two web apps with an AI agent
- The MCP server
- How the tools are shaped, and why
- Playwright MCP vs the Playwright CLI: which fits when
- Playwright MCP: browser is already in use, and the fix
- Playwright MCP best practices: four decisions that matter
- Playwright MCP with a proxy, and the three leaks it leaves
- A browser MCP server in GitHub Copilot: setup and limits
- Using a browser MCP server for web scraping: the pattern
- Which LLM for browser automation: the four properties
- How to build a browser agent, and what to take instead
- Getting an AI agent to log into a website: three routes
- MCP tools, resources and prompts: who controls each
- How many MCP tools is too many? The context arithmetic
- How to build an MCP server: the decisions, not the scaffold
- Local or remote MCP server: what changes, and what does not
- Writing an MCP client in Python: the thirty-line version