AI agents and MCP servers you can connect today — free, with install steps

Two kinds of AI agent — and which one you need

If your AI needs to see or change something outside the chat, you want an MCP server. If it needs to hand a focused job to a specialist without cluttering your session, you want a subagent. Need the steps for a job rather than a tool? That’s a skill.

What is an MCP server?

An MCP server is a small program that gives an AI a set of tools through the Model Context Protocol, an open standard. The AI app (the client) asks the server what tools it has, then calls them when a request needs one. Because the protocol is shared, the same server works in Claude, Claude Code, Cursor and every other client that supports MCP.

There are two ways to run one. Remote servers are hosted by the vendor — you paste an address and sign in; nothing is installed. Local servers run on your computer through npx, uvx or Docker, and can reach your files and local databases.

How to connect an MCP server

Claude (web and desktop)

Go to Customize → Connectors, press + → Add custom connector, paste the server address and press Add. If the server uses OAuth, its client ID and secret go under Advanced settings.

Claude Code

# hosted server
claude mcp add --transport http <name> <url>

# server that runs on your machine
claude mcp add <name> -- npx -y <package>

Cursor

// .cursor/mcp.json (one project) or ~/.cursor/mcp.json (everywhere)
{
  "mcpServers": {
    "<name>": { "command": "npx", "args": ["-y", "<package>"] }
  }
}

For the servers we have checked by hand, the agent page fills in the real package, address and any key it needs; for the rest it points you to the author’s README rather than guess. The connect guide walks through it with screenshots, and you can even put AgentAlley itself inside your AI (a paid MCP plan) to search this catalogue from a chat.

Best MCP servers to start with

Chosen for being useful on day one, maintained, and — where one exists — the vendor’s own official server. Connect two or three you’ll really use; each extra server is more tools for the model to choose between.

1. 1Context7Pulls current, version-specific library docs into the conversation, so Claude stops writing code against APIs that have changed. The one most people install first.by upstash · MIT · full source on the page
2. 2Playwright MCPMicrosoft's browser server: Claude opens pages, clicks, fills forms and reads what is on screen. Good for testing your own site end to end.by microsoft · Apache-2.0 · full source on the page
3. 3Github MCP serverGitHub's official server: read repos, open and review pull requests, manage issues and Actions without leaving the chat.by github · MIT · full source on the page
4. 4Firecrawl MCP ServerScrape, crawl and search the web and get clean Markdown back. Has a hosted endpoint you can try without a key.by firecrawl · MIT · full source on the page
5. 5Filesystem MCP ServerLets Claude read and write files inside folders you allow. Point it at one project folder, never your whole drive.by modelcontextprotocol · MIT · full source on the page
6. 6Fetch MCP ServerFetches a web page and turns it into Markdown the model can read. The simplest way to give Claude a live URL.by modelcontextprotocol · MIT · full source on the page
7. 7Notion MCP serverNotion's official server: search, read and edit pages and databases in your workspace.by makenotion · MIT · full source on the page
8. 8Supabase MCP ServerQuery tables, run migrations, read logs and manage branches in your Supabase projects.by supabase-community · Apache-2.0 · full source on the page
9. 9Chrome devtools MCPClaude drives a real Chrome with DevTools open: performance traces, network requests, console errors.by ChromeDevTools · Apache-2.0 · full source on the page
10. 10Exa MCP ServerHosted web search and page fetching built for AI agents, with nothing to install.by exa-labs · MIT · full source on the page
11. 11Postgres MCP ProPostgres health checks, index tuning and explain plans — with a restricted read-only mode worth turning on.by crystaldba · MIT · full source on the page
12. 12Sequential Thinking MCP ServerA scratchpad tool that makes the model reason through multi-step problems one explicit step at a time.by modelcontextprotocol · MIT · full source on the page

All engineering agents →

Claude Code subagents worth installing

A subagent is a Markdown file that defines a specialist: a name, a description of when Claude should hand work to it, which tools it may use, and its instructions. Drop the file in .claude/agents/ for one project or ~/.claude/agents/ for all of them, or manage them with the /agents command. Each one works in its own context window, so a long review doesn’t fill up your main session.

---
name: code-reviewer
description: Use after writing or changing code, before committing.
tools: Read, Grep, Glob
---

You review diffs for bugs, security holes and unclear names.
Report problems ordered by severity. Do not rewrite the code.

1. 1Code reviewerA reviewer subagent that checks code for correctness, security and maintainability before you merge.by VoltAgent · MIT · full source on the page
2. 2DebuggerDiagnoses failures and traces them to a root cause, in its own context so your main session stays clean.by VoltAgent · MIT · full source on the page
3. 3Architect reviewerEvaluates system design decisions: boundaries, coupling, scalability and the trade-offs you are making.by VoltAgent · MIT · full source on the page
4. 4Agent organizerBreaks a large task into subtasks and matches each one to the right specialist agent.by VoltAgent · MIT · full source on the page
5. 5Accessibility testerChecks interfaces against WCAG and reports what blocks keyboard and screen-reader users.by VoltAgent · MIT · full source on the page
6. 6🧭 Product Manager AgentA product lead persona: discovery, strategy, prioritisation and roadmap, end to end.by msitarzewski · MIT · full source on the page
7. 7SEO SpecialistTechnical SEO, content optimisation and link strategy in one specialist role.by msitarzewski · MIT · full source on the page
8. 8📊 Financial Analyst AgentFinancial modelling, forecasting and scenario analysis for plans and budgets.by msitarzewski · MIT · full source on the page

All 171 subagents and roles →

Browse agents by job

Using agents safely

• Prefer official servers. GitHub, Notion, Supabase, Cloudflare and others publish their own — use those over a stranger’s wrapper.
• Start read-only. Many servers have a read-only or restricted mode. Turn write access on only when you need it.
• Scope local access. A filesystem server should see one project folder, not your home directory.
• Keep approvals on. Choose “Allow once” until you trust a tool.
• Keys go in headers or env vars, never pasted into a chat message.
• Read before you run. Every listing here shows the full source, author and licence, with a link to the exact file on GitHub.

AI agents & MCP FAQ

Keep going