Guide
What is Model Context Protocol (MCP)?
An open standard for connecting AI clients (Claude Desktop, Cursor, Zed, Continue) to external tools and data sources. Write the server once — every MCP-aware client can use it.
The 60-second pitch
Anthropic published the MCP spec in November 2024. The problem it solved: every AI chat client was re-implementing the same integrations (filesystem, GitHub, Slack, web search, Postgres…) and every integration had to be re-built for every new client. MCP turns that N×M matrix into N+M — one server per tool, one client per app.
Today MCP is supported by Claude Desktop, Cursor, Zed, Continue, Cody, Goose, Sourcegraph, and a growing list of agent frameworks. Anthropic still maintains the spec but doesn't gate participation — anyone can publish an MCP server, and the protocol is intentionally simple (stdio + JSON-RPC).
What can MCP servers do?
Anything you can wrap in a tool call. The reference servers Anthropic ships cover the common 80%:
- ›Filesystem — read/write files in a sandboxed directory
- ›GitHub — issues, PRs, code reads, workflow runs
- ›Databases — Postgres, SQLite, MongoDB, Redis, vector stores (Qdrant, Pinecone, Weaviate)
- ›Search — Brave, Tavily, Exa, Perplexity, Elasticsearch
- ›Browser automation — Puppeteer, Playwright
- ›Communication — Slack, Discord, Telegram, email APIs
- ›CRM / business — HubSpot, Salesforce, Linear, Atlassian, Stripe
- ›Cloud / infra — AWS, Cloudflare, Vercel, Docker, Kubernetes
Plus your own — the spec is ~200 lines of TypeScript and the SDKs (TypeScript, Python, Rust) make publishing a new server a one-day project.
Install your first MCP server in 5 minutes
- Pick a server from the /mcp directory. Start with
anthropic-filesystemfor a safe sandbox. - Open
claude_desktop_config.json. Mac:~/Library/Application Support/Claude/. Windows:%APPDATA%\Claude\. Create the file if it doesn't exist. - Paste the config snippet from the Prismix detail page. Each server detail renders a ready-to-paste
{ "mcpServers": { ... } }block — replace any${ENV_VAR}placeholders with your own values. - Restart Claude Desktop fully (quit + reopen, not just close the window).
- Verify in chat — ask Claude "what tools do you have?". The new server should appear with its tool names. To install a whole stack at once, browse /mcp/bundles — every public bundle has a merged config snippet that installs every server in one paste.
The MCP ecosystem map
Top servers by category. Click a category for the full list.
Data
19 serversFilesystem, databases, object storage, knowledge bases
Top: Filesystem · Postgres · SQLite
Dev tools
12 serversGitHub, GitLab, issue trackers, observability, debugging
Top: GitHub · GitLab · Sentry
AI
9 serversSearch engines, image gen, model inference, embeddings
Top: Brave Search · Memory · Sequential Thinking
Web
4 serversHTTP fetch, browser automation, scraping, screenshots
Top: Fetch · Puppeteer · Browser (Stealth)
Communication
6 serversSlack, Discord, Telegram, email, SMS
Top: Slack · Discord · Telegram
Productivity
13 serversNotion, Obsidian, todo lists, calendars, time tracking
Top: Time · Google Maps · Linear
System
5 serversShell, SSH, Docker, Kubernetes, process management
Top: Shell · Docker · Kubernetes
Media
8 serversSpotify, YouTube, audio + video editing, transcription
Top: EverArt · OpenAI Images · Replicate
Finance
4 serversStripe, Plaid, accounting, market data
Top: Stripe · Plaid · Alpaca Trading
FAQ
What is Model Context Protocol?
Model Context Protocol (MCP) is an open standard introduced by Anthropic in late 2024 for connecting LLM clients (Claude Desktop, Cursor, Zed, Continue) to external data sources and tools. Instead of every chat app re-implementing every integration, MCP defines a single server-client protocol — write the server once, every MCP-aware client can use it.
Who created MCP?
Anthropic published the MCP specification + reference implementations in November 2024. The spec is open and the reference servers ship under MIT-style licenses. Anthropic maintains the protocol but doesn't gate participation — anyone can publish an MCP server.
Which AI clients support MCP?
Claude Desktop (Anthropic), Cursor, Zed, Continue, Cody, Goose, and Sourcegraph all ship MCP integrations as of mid-2026. Most read a claude_desktop_config.json or equivalent JSON config file declaring which servers to run.
What can MCP servers do?
Anything you can wrap in a tool call — read files, query databases (Postgres, SQLite, vector stores), search the web (Brave, Tavily, Exa, Perplexity), drive a browser (Puppeteer, Playwright), call third-party APIs (GitHub, Slack, Linear, Stripe, HubSpot), or expose your own internal tools. MCP servers are typically small npm / uvx packages.
Is MCP only for Claude?
No. While Anthropic authored the protocol, MCP is an open standard and every major code-assistant client now supports it. The same Postgres MCP server runs identically under Cursor, Claude Desktop, or Zed.
How do I find good MCP servers?
Prismix's MCP directory at /mcp tracks 80+ curated MCP servers plus ~500 auto-discovered from GitHub topic search. Each entry has install commands, Claude Desktop config snippets, GitHub stats, community ratings, and a per-release timeline. Bundle them at /mcp/bundles to install a whole stack at once.
Ready to dig in?
The Prismix MCP directory has 80+ curated servers plus ~500 auto-discovered from GitHub. Bundle them into named stacks, get release notifications, install whole workflows with one paste.