Using Existing MCP Servers

Hundreds of MCP servers exist — filesystem, GitHub, databases. Using them well is a security and governance exercise wearing a configuration file.

▶ Watch this reel

What you'll learn

The reference server zoo
Connecting from hosts
Trust: the supply-chain question
MCP gateways

Remember this

Consume before building: reference servers (filesystem, DB, GitHub) are maintained and spec-current — fork and scope, don't start from zero
Host configs are security policy: the filesystem path argument is the sandbox, tokens deserve fine-grained scopes, remote URLs need an operator you trust
Servers are supply-chain dependencies: provenance, pinned versions, audited tool surfaces; gateways centralize auth, audit, and policy at enterprise scale

Reference servers

Filesystem (scoped dirs) · databases (read-only modes) · GitHub · fetch · hundreds community-maintained.
Strategy: consume → fork/scope → build only what's missing.

Connecting

stdio = local config: command + args + env. Config IS security policy.
Filesystem path arg = the sandbox. Tokens fine-grained. Remote URLs need trusted operators.

Trust (supply chain)

Provenance (official/vendor) · read the tool surface · minimum scopes · pin versions · track spec updates · registry (no shadow IT).

Gateways

Enterprise control plane: central auth (SSO), full audit, policy, server registry.
Credentials off endpoints. Critical infra — but governance must stay fast.

Code: The server-vetting checklist

# Before adding ANY MCP server to your stack:

provenance:
  □ Official/reference impl, or vendor-authored for its own API?
  □ Known maintainers, public repo, issue responsiveness?

surface:
  □ Read the exposed tools — does each match the server's stated job?
  □ Any unexpected capabilities (shell, network, file writes)?

permissions:
  □ Minimum scopes? Read-only where possible?
  □ Credentials scoped to specific repos/paths, not broad?

operations:
  □ Version pinned (npx pkg@1.2.3 / lockfile), not latest?
  □ Tracks MCP spec revisions (post-2026-07-28 aware)?
  □ Logged in the team server registry (no shadow IT)?

# 2 minutes per server. Cheaper than one incident.