There are 91,000+ Claude skills and 50,000+ MCP listings out there. You need about a dozen. This is the short list, kept deliberately small: five MCP configs worth wiring up, seven skills worth installing, and just as usefully, what to skip. Copy-paste blocks included, checked against the real package names so they work first time.
MCP servers are what let Claude actually touch things: your files, the web, a database, your repo. These five are the official reference servers, run locally over stdio, no accounts or API keys to wrangle for most of them. Paste each block into your MCP config (in Claude Desktop that is claude_desktop_config.json; in Claude Code, claude mcp add or your project's .mcp.json) under "mcpServers", swap in your own paths, and restart.
Why: the single most useful MCP there is. It gives Claude read and write access to folders you name, nothing else. Everything else on this list is a nice-to-have; this one is the reason "Claude in a folder" works at all.
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/path/to/your/folder"
]
}
}
}
Only list folders you are genuinely fine with Claude reading and editing. It runs with your own file permissions.
Why: lets Claude pull a real webpage and read it as clean text instead of guessing from training data. The difference between "probably right" and "just checked."
{
"mcpServers": {
"fetch": {
"command": "uvx",
"args": ["mcp-server-fetch"]
}
}
}
Needs uv installed (uvx ships with it). On Windows, if it hangs, add "env": {"PYTHONIOENCODING": "utf-8"} to the block.
Why: Claude can read history, diff, and status on a real repo instead of you pasting code back and forth. Small thing, saves a lot of copy-paste.
{
"mcpServers": {
"git": {
"command": "uvx",
"args": ["mcp-server-git", "--repository", "/path/to/your/repo"]
}
}
}
Read and inspect are safe defaults. Treat anything that commits or pushes as a step you approve, not one you automate blind.
Why: a small local knowledge graph so Claude can remember facts about your business across sessions instead of you re-explaining it every time. This is the one that makes Claude feel like it knows you.
{
"mcpServers": {
"memory": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-memory"],
"env": {
"MEMORY_FILE_PATH": "/path/to/your/memory.jsonl"
}
}
}
}
It is a plain file on your machine. Back it up like you would any other business record.
Why: forces Claude to lay out its reasoning step by step before answering, instead of jumping straight to a confident-sounding conclusion. Worth having wired up for anything with real stakes: pricing, contracts, numbers.
{
"mcpServers": {
"sequential-thinking": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
}
}
}
On Windows, if npx will not launch directly from the config, wrap it: "command": "cmd", "args": ["/c", "npx", "-y", "@modelcontextprotocol/server-sequential-thinking"].
Skipped on purpose: the official SQLite reference server. It has an unpatched SQL injection issue and has not been updated since it was archived. If you need a database MCP, wait for a maintained one or ask us and we will point you at the current pick.
Skills are the other half: they tell Claude how to do a job well, not just what tools it can reach. These seven are general enough to earn a permanent spot, not novelty extensions you try once and forget.
Most of the 91,000 fall into a handful of categories that sound useful and mostly are not, for a small business owner running Claude day to day.
Want this list, plus two free tool packs built for Australian small business? Grab the packs on the home page, no catch.
Get the packs and the guide