@modelcontextprotocol/server-filesystem — read/write local files. Good baseline for tool design (clear input schemas, scoped paths).
MCP examples
MCP server examples.
A curated list of MCP servers shipping in 2026, what they do, how they're configured, and the patterns worth lifting. All examples have public source — read the code before copying.
Completely free for 30 days. No credit card required.
01Teams
Reference servers from the MCP team
@modelcontextprotocol/server-fetch — fetch a URL and return its content. Cleanest example of a stateless tool with structured returns.
@modelcontextprotocol/server-memory — knowledge-graph memory the agent can read + write. Demonstrates resource exposure alongside tools.
@modelcontextprotocol/server-git — repo introspection (status, log, diff, blame). Shows how to handle filesystem-anchored state inside a tool.
@modelcontextprotocol/server-everything — every primitive the SDK exposes (tools, resources, prompts, sampling) in one place. The canonical end-to-end reference.
All open source at github.com/modelcontextprotocol/servers — read these first if you're learning the SDK. Older server-github / -postgres / -brave-search were archived in 2025; vendor-maintained replacements exist (e.g. github.com/github/github-mcp-server, github.com/stripe/agent-toolkit).
02Details
Browser automation
@playwright/mcp — Microsoft's official MCP server wrapping Playwright. Tools for navigate / screenshot / click / type. Stdio + HTTP modes.
sendblue-browser-mcp — wraps a stealth-patched Chromium daemon. Tools for navigate / screenshot / script eval / CDP attach. Useful when you need bot-resistant browsing.
Both ship as npm packages, install via npx -y <pkg>. Browser MCP servers are popular because agents need to test their own UI flows.
03Details
API + SaaS wrappers
Stripe (github.com/stripe/agent-toolkit), GitHub (github.com/github/github-mcp-server), Linear (community linear-mcp), plus Atlassian, Notion, Slack — official or strong community servers wrapping each provider's REST API.
Pattern: one tool per API operation, auth via env-supplied API key, structured returns matching the underlying response schema.
Most of these are hand-written today. Teams shipping more than 5-10 endpoints increasingly generate from OpenAPI to avoid the maintenance tax.
04Details
Patterns worth copying
Sanitise error messages — Playwright MCP and sendblue-browser-mcp both truncate Playwright's verbose error output to one line before returning. Avoids leaking internal IDs.
Allowlist URL schemes for any tool that takes a URL input — block file://, chrome://, etc. (the sendblue-browser-mcp navigate tool does this).
Cap input sizes — script.code capped at 200kB, cookies capped at 500 entries. Prevents agents from accidentally OOMing your server with a runaway loop.
Image returns as base64 with explicit mimeType — every screenshot-returning server uses this pattern.
05Details
Anti-patterns to avoid
Accepting API keys as a tool input parameter — agents will paste them into chat history. Always read auth from env.
Exposing every internal endpoint as a tool — pick the operations agents actually need; hide the rest.
Long tool descriptions — agents weigh tool selection by description length / clarity. Aim for one clear sentence.
Synchronous network calls that block the event loop — use async/await everywhere (Node) or asyncio (Python).
06Outputs
Generated from OpenAPI (preview)
MCP-server emission is in preview alongside Bloom's TypeScript and Python SDK pipeline — same OpenAPI input, same dashboard, with the patterns above baked in as they land.
Target output: one tool per operation, auth from env, error sanitisation, size caps, base64 image handling, sanitised allow-list for URL inputs.
Same workflow as Bloom's TypeScript/Python SDK generation (which ships today). Free for 30 days; no credit card.