Where do you find MCP servers, and which can you trust?

Short answer

Find MCP servers in the official MCP Registry (in preview), on the vendor's own site or GitHub, and in community lists such as awesome-mcp-servers. None of these vouch for safety: the registry checks who published a server, not what its code does. Prefer the vendor's official server, check the owner, transport, sign-in and write tools, and test it before the team uses it.

Four places to look, and what each one promises

Source What it is What it checks What it does not check
MCP Registry The official metadata registry, backed by Anthropic, GitHub, PulseMCP and Microsoft, among others. In preview. That the publisher owns the namespace, via GitHub, DNS or HTTP The server’s code; it leaves security scanning to npm, PyPI, Docker Hub and downstream marketplaces
Reference servers Seven servers in modelcontextprotocol/servers, maintained by the MCP steering group Maintained by the protocol’s own people Production readiness; the project calls them educational examples
The vendor’s own server For example GitHub’s official server, or Sentry’s remote server The vendor stands behind it Whether the scopes fit your team
Community lists awesome-mcp-servers on GitHub, and directories such as Glama Nothing formal; anyone can propose an entry Ownership, code, maintenance

The MCP Registry

The registry stores a server.json per server: a unique name such as io.github.user/server-name, where the package or remote URL lives, and how to run it. Names follow reverse DNS, so only the owner of a GitHub account or a domain can publish under that name. That stops impersonation in the registry, not bad code.

Two limits matter for companies. The registry does not take private servers: a server on mcp.acme-corp.internal or on a private package registry does not belong there, and the project recommends running your own private registry for those. And the registry is meant to be read by marketplaces and aggregators, not directly by your AI app.

Reference servers

The modelcontextprotocol/servers repository now holds only Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. Servers that once lived there, including GitHub, GitLab, Google Drive, PostgreSQL, Slack and SQLite, were moved to an archive. If a tutorial tells you to install one of those old packages, look for the vendor’s current server instead.

Community lists

“Awesome” lists are useful for discovery. They show breadth: database servers, browser automation, design tools, internal-tool connectors. A spot on a list is not a review, though, so treat each entry as a lead to check.

The common misunderstanding: listed means vetted

The most common mistake is to treat “it is in the registry” or “it has many GitHub stars” as a security check. The registry’s own documentation says it focuses on namespace authentication and metadata, and relies on the broader ecosystem for scanning code. Stars measure popularity.

The MCP security guidance explains why that matters. A local server is a program that runs with your own user rights. The examples it gives are a start command in a config that also sends your SSH key somewhere, a malicious package, and an insecure local server reached through DNS rebinding. A remote server sees everything you send it and holds a token to your systems.

How to vet a server before the team uses it

  1. Owner. Is it published by the vendor of the system it connects to, under the vendor’s GitHub organization or domain? If not, who maintains it, and how recently?
  2. Transport and sign-in. Remote servers should use OAuth so each person signs in as themselves; the specification forbids a server from passing your token on to another API. Local servers should get credentials from environment variables, not from a key pasted into a shared file.
  3. Scopes and write tools. List the tools. Can it delete or send? Prefer servers that let you narrow this; GitHub’s server, for example, has --toolsets to pick tool groups and --read-only to switch off write tools.
  4. The exact start command. For a local server, read the full command and arguments before you approve it. The MCP guidance says clients that offer one-click setup must show it in full.
  5. Test it alone first. Run it in one person’s setup, look at what each tool returns, then roll it out.
  6. Control the list. In Claude Code, a project .mcp.json shares an approved set through git, and Claude Code asks for approval before using project servers. Administrators can go further with managed MCP configuration (managed-mcp.json, allowedMcpServers, deniedMcpServers).

For broader company rules, see is MCP safe to use in a company. If no trustworthy server exists for your system, building your own is often less work than expected.

What this means for a team

Without an agreed list, every person picks servers on their own, and the same system gets connected three different ways with three different permission sets. Decide per system which server is approved, whether it runs local or remote, and who may add new ones. The difference between the two is explained in local vs remote MCP servers.

Where Bombos fits

Bombos is a plugin that gives every AI tool your team uses (Claude, Claude Code, ChatGPT, Copilot, Grok, Codex, Cursor, your own agents) one shared memory, so the team’s knowledge and learned rules do not depend on which servers each person happened to install. It reads what already sits in mail, documents and repos, shows what each answer rests on, and runs in our cloud or on your own server. One person can try it first with their own agents.

Related questions

Is there an official list of MCP servers?

Yes, the MCP Registry at registry.modelcontextprotocol.io. It is in preview, holds metadata that points to packages or remote URLs, and verifies that a publisher owns its namespace. It does not scan server code.

Are the reference servers from the MCP project safe for production?

The project says no: its reference servers (Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time) are educational examples, not production-ready. Older ones such as GitHub, Slack and PostgreSQL were archived.

What is the best MCP server?

There is no single best one. The useful ones are the official servers of tools your team already uses, connected with only the permissions the work needs.

Can we limit which MCP servers employees use?

In Claude Code, yes: managed MCP configuration can deploy a fixed set with managed-mcp.json and allow or block servers with allowedMcpServers and deniedMcpServers. Other AI tools have their own admin settings.

Sources

  1. The MCP Registry (Model Context Protocol)
  2. Model Context Protocol servers (reference implementations)
  3. awesome-mcp-servers
  4. Security Best Practices, MCP specification 2026-07-28
  5. GitHub MCP Server
  6. Connect Claude Code to tools via MCP
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.