MCP Server Catalog
A categorized list of MCP servers that are maintained as of November 2026: what replaced the archived packages, local (stdio) and remote (Streamable HTTP) setup, Claude Code and Claude Desktop configuration, and safe use.
Updated:
What is MCP, and where do I start?
MCP (Model Context Protocol) is an open standard for AI agents to connect to external tools and data. Anthropic announced it in November 2024; since 9 December 2025 it has been governed by the Agentic AI Foundation (AAIF) under the Linux Foundation. An MCP server exposes a service to Claude as a "tool". Claude calls it, the result enters context and informs the next reply.
As of November 2026 the current specification revision is 2026-07-28. It makes the protocol stateless: no initialize handshake and no sessions. The old HTTP+SSE transport is formally Deprecated; remote servers use Streamable HTTP. In practice, most large vendors now run their own remote MCP server, and you just add a URL and sign in with OAuth.
This guide groups servers that are maintained as of November 2026 by category. For each: what it does, when to pick it, setup, the main pitfall, and its source. We checked package names and release dates against npm, PyPI and the vendor's GitHub repository.
Before you install: watch out for archived servers
Many examples online, including the previous version of this catalog, point at packages that are no longer maintained. The modelcontextprotocol/servers repository now hosts only seven reference servers (Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time); the rest moved to modelcontextprotocol/servers-archived.
Status on npm as of November 2026:
| Package | Status | Replacement |
|---|---|---|
@modelcontextprotocol/server-github |
Deprecated: "Archived and unmaintained; no security guarantees" | github/github-mcp-server |
@modelcontextprotocol/server-postgres |
Deprecated: same notice | MCP Toolbox for Databases |
@modelcontextprotocol/server-puppeteer |
Deprecated: same notice | @playwright/mcp |
@modelcontextprotocol/server-brave-search |
Deprecated: "Package no longer supported" | @brave/brave-search-mcp-server |
server-slack, server-gdrive, server-google-maps, server-redis, server-everart, server-aws-kb-retrieval |
Deprecated: "Package no longer supported" | See the relevant sections below |
server-sentry, server-fetch, server-git, server-time, server-sqlite (npm) |
Never existed on npm | Sentry's own server; the mcp-server-* packages on PyPI for fetch/git/time |
@stripe/mcp |
Deprecated: "No longer maintained" | Stripe's remote server https://mcp.stripe.com |
Before you install a server, check two things:
npm view <package> deprecated # empty output = not deprecated
npm view <package> time.modifiedFor PyPI packages, check the latest release date at https://pypi.org/pypi/<package>/json; for a GitHub repo, look for the "This repository has been archived" banner.
Official MCP Registry: registry.modelcontextprotocol.io is the official catalog where publishers register their servers. As of November 2026 its repository still describes it as a preview and says the API is frozen at v0.1. The Registry is a place to look, not a security review: verify that the publisher really is the company it claims to be.
Anatomy of a config
There are two kinds of server:
- Local (stdio): the server runs as a process on your machine. It takes three things: command (how to run it), args (parameters), env (secrets). Most start with
npx,uvxordocker run. - Remote (Streamable HTTP): the vendor runs the server; you add a URL and usually authenticate with OAuth in the browser.
Claude Code
The recommended way to add a server is claude mcp add:
# Remote server
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
# Local stdio server (the command to run goes after --)
claude mcp add --transport stdio git -- uvx mcp-server-git --repository /path/to/repo
# Choosing a scope
claude mcp add --scope project --transport http notion https://mcp.notion.com/mcp--scope takes three values:
| Scope | Loads in | Stored in |
|---|---|---|
local (default) |
This project only, just for you | ~/.claude.json |
project |
This project, shared with the team via Git | .mcp.json at the project root |
user |
All your projects, just for you | ~/.claude.json |
For remote servers that need OAuth, sign in from inside Claude Code with the /mcp command. You can also write .mcp.json by hand:
{
"mcpServers": {
"sentry": {
"type": "http",
"url": "https://mcp.sentry.dev/mcp"
},
"git": {
"command": "uvx",
"args": ["mcp-server-git", "--repository", "${REPO_PATH:-.}"]
}
}
}A remote entry needs "type": "http"; Claude Code treats an entry with a url but no type as stdio and skips it. ${VAR} and ${VAR:-default} expansion is a Claude Code feature for .mcp.json: it works in command, args, env, url and headers. Not every client supports it, so check the other client's docs before you copy a config across.
Claude Desktop
Local stdio servers go in claude_desktop_config.json:
- macOS:
~/Library/Application Support/Claude/claude_desktop_config.json - Windows:
%APPDATA%\Claude\claude_desktop_config.json - Linux:
~/.config/Claude/claude_desktop_config.json
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/me/projects"]
}
}
}Remote servers don't go in this file; you add them as custom connectors. On Pro and Max plans, go to Customize > Connectors, click "+ Add", then "Add custom connector", enter a name and the remote MCP server URL, and choose the authentication settings. On Team and Enterprise plans an Owner first adds it under Organization settings > Connectors.
Don't write tokens into a file that goes into Git. In Claude Code, read them from environment variables with ${VAR}; the Claude Desktop file stays on your machine, but don't share it either.
Development environment
filesystem
Local read/write access to folders. Claude opens and edits files directly. Most useful in Claude Desktop; Claude Code already has its own file tools.
"filesystem": {
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"/Users/me/projects",
"/Users/me/Documents/notes"
]
}Pitfall: never set the root to / or your home folder; that hands Claude every file. Restrict it to the subfolders you actually work in.
Source: npm @modelcontextprotocol/server-filesystem 2.2.0 (4 November 2026), github.com/modelcontextprotocol/servers/tree/main/src/filesystem.
git
Read, search and operate on a local Git repository (status, diff, log, commit, branch). A Python package, run with uvx.
"git": {
"command": "uvx",
"args": ["mcp-server-git", "--repository", "/path/to/repo"]
}Pitfall: the commit and branch tools are enabled too. If you only want to read, have your client ask for approval on write tools.
Source: PyPI mcp-server-git 2026.11.4 (4 November 2026), github.com/modelcontextprotocol/servers/tree/main/src/git.
github
Issues, PRs, Actions, code search, security alerts. GitHub's own server replaced the old @modelcontextprotocol/server-github. There are two ways to run it:
Remote (recommended, nothing to install): https://api.githubcopilot.com/mcp/. Append /readonly to the URL for a read-only version.
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/readonly",
"headers": { "Authorization": "Bearer ${GITHUB_PAT}" }
}Local (Docker): if you use GitHub Enterprise Server, there is no remote server, so use this path.
claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=$GITHUB_PAT -e GITHUB_READ_ONLY=1 \
-- docker run -i --rm -e GITHUB_PERSONAL_ACCESS_TOKEN -e GITHUB_READ_ONLY ghcr.io/github/github-mcp-serverPitfall: scope the token tightly: a fine-grained token, only the repos you need, only the permissions you need. Narrowing tools with GITHUB_TOOLSETS (e.g. repos,issues) cuts both the permissions in play and context usage. Issues and comments on public repos are untrusted text entering the model and can carry prompt injection.
Source: github.com/github/github-mcp-server, latest release v2.1.0 (3 November 2026).
sentry
Query production errors: "top 5 errors in the last 24h", "show me this issue's stack trace". There was never an @modelcontextprotocol/server-sentry on npm; Sentry ships its own server.
Remote (recommended): https://mcp.sentry.dev/mcp, sign in with OAuth.
claude mcp add --transport http sentry https://mcp.sentry.dev/mcpLocal stdio (for self-hosted Sentry): @sentry/mcp-server. Sign in once with npx @sentry/mcp-server@latest auth login or set SENTRY_AUTH_TOKEN; set SENTRY_HOST for your own instance.
"sentry": {
"command": "npx",
"args": ["-y", "@sentry/mcp-server@latest"],
"env": { "SENTRY_AUTH_TOKEN": "${SENTRY_AUTH_TOKEN}", "SENTRY_HOST": "sentry.example.com" }
}Pitfall: error messages and breadcrumbs can contain user input; treat them as untrusted text entering the model.
Source: npm @sentry/mcp-server 0.49.0 (7 November 2026), github.com/getsentry/toolkit.
Databases
postgres
Read schema and run SQL. The maintained replacement for the archived @modelcontextprotocol/server-postgres is Google's MCP Toolbox for Databases. --prebuilt=postgres loads ready-made tools (list_tables, execute_sql and others); the connection is read from environment variables.
"postgres": {
"command": "npx",
"args": ["-y", "@toolbox-sdk/server", "--prebuilt=postgres", "--stdio"],
"env": {
"POSTGRES_HOST": "localhost",
"POSTGRES_PORT": "5432",
"POSTGRES_DATABASE": "mydb",
"POSTGRES_USER": "claude_ro",
"POSTGRES_PASSWORD": "${POSTGRES_PASSWORD}"
}
}Pitfall: execute_sql runs whatever raw SQL the model writes. Toolbox's own docs say it plainly: tool annotations and client confirmations are not a security boundary; database permissions are. Create a SELECT-only user:
CREATE ROLE claude_ro WITH LOGIN PASSWORD '...';
GRANT CONNECT ON DATABASE mydb TO claude_ro;
GRANT USAGE ON SCHEMA public TO claude_ro;
GRANT SELECT ON ALL TABLES IN SCHEMA public TO claude_ro;
ALTER ROLE claude_ro SET default_transaction_read_only = on;Pointing it at a replica or a copy instead of the production database is better still.
Source: npm @toolbox-sdk/server 1.16.0 (5 November 2026), github.com/googleapis/mcp-toolbox. The same tool supports MySQL (--prebuilt=mysql), SQL Server, Oracle, BigQuery and other engines.
sqlite
Local dev databases, small analytics datasets. The old mcp-server-sqlite (PyPI) is archived and its last release shipped in April 2025; use MCP Toolbox here too:
"sqlite": {
"command": "npx",
"args": ["-y", "@toolbox-sdk/server", "--prebuilt=sqlite", "--stdio"],
"env": { "SQLITE_DATABASE": "/path/to/db.sqlite" }
}Pitfall: work on a copy of the file; execute_sql can write too.
Source: github.com/googleapis/mcp-toolbox, internal/prebuiltconfigs/tools/sqlite.yaml.
mongodb
MongoDB's own server. Every example in its docs ships with --readOnly; keep it that way.
"mongodb": {
"command": "npx",
"args": ["-y", "mongodb-mcp-server@latest", "--readOnly"],
"env": { "MDB_MCP_CONNECTION_STRING": "${MDB_MCP_CONNECTION_STRING}" }
}Pitfall: --readOnly is a server-side guard; also give the user in the connection string a read-only role in the database.
Source: npm mongodb-mcp-server 3.2.2 (29 October 2026), github.com/mongodb-js/mongodb-mcp-server.
redis
Redis's own server (Python, uvx) replaces the archived @modelcontextprotocol/server-redis:
"redis": {
"command": "uvx",
"args": ["--from", "redis-mcp-server@latest", "redis-mcp-server", "--url", "redis://localhost:6379/0"]
}Pitfall: the server exposes write and delete commands too. Define a read-only user with Redis ACLs and use it in the URL.
Source: PyPI redis-mcp-server 0.5.1 (5 August 2026), github.com/redis/mcp-redis.
Communication and work tools
slack
Search channels and messages, send messages. Slack now runs its own remote server in place of the archived @modelcontextprotocol/server-slack: https://mcp.slack.com/mcp. According to Slack's docs it is not open to every MCP client: the client must be backed by a registered Slack app, which the workspace admin approves. Claude.ai and Claude Code are among the listed ready-made clients.
"slack": {
"type": "http",
"url": "https://mcp.slack.com/mcp"
}Review the permissions on Slack's consent screen during authorization; start with read-only access where you can.
Pitfall: the moment you grant chat:write, one mistake can post to the wrong channel. Add write scopes only when you really need them.
Source: docs.slack.dev/ai/slack-mcp-server/, code.claude.com/docs/en/mcp.
google drive / gmail
As of November 2026 Google has no official general-purpose MCP server for Drive. The Workspace server on Google's own list (github.com/google/mcp) ships as a Gemini CLI extension. Don't use the archived @modelcontextprotocol/server-gdrive.
In Claude.ai and Claude Desktop the practical route is Anthropic's built-in Google Workspace connectors: Gmail, Google Calendar and Google Drive. You sign in directly with your Google account and Claude mirrors your existing Google permissions. On Team and Enterprise plans an Owner has to enable them at the account level first.
Pitfall: emails and shared documents are outside text entering the model. "What does this email say?" is a prompt injection doorway.
Source: support.claude.com/en/articles/10166901-use-google-workspace-connectors.
linear / notion / jira
Issue-tracker integration: "list my open tickets", "comment on this". All three are vendor-run remote servers:
claude mcp add --transport http linear https://mcp.linear.app/mcp
claude mcp add --transport http notion https://mcp.notion.com/mcp
claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp- Linear: OAuth; you can also pass an API key in an
Authorization: Bearerheader, including a restricted key for read-only access. Source:linear.app/docs/mcp. - Notion: Notion says the local
@notionhq/notion-mcp-serverpackage (npm 2.6.1, 29 October 2026) is not its priority and may be sunset; it recommends the remote server. Source:developers.notion.com/docs/mcp. - Atlassian (Jira, Confluence, Bitbucket and more): OAuth 2.1 or API token. The v2 URL is recommended for new setups. Source:
github.com/atlassian/atlassian-mcp-server.
Pitfall: these servers act with your account's permissions. Assume Claude can close any issue you can; require approval for write tools.
asana
Query tasks and projects, create tasks. Asana's remote server is https://mcp.asana.com/v2/mcp and requires OAuth and Streamable HTTP. On Enterprise+ tiers, admins can allow or block MCP clients through Asana's app management.
Source: developers.asana.com/docs/using-asanas-mcp-server.
Cloud and SaaS
stripe
Look up customers, payments and subscriptions. The @stripe/mcp npm package is deprecated with "No longer maintained"; Stripe points to its remote server: https://mcp.stripe.com.
claude mcp add --transport http stripe https://mcp.stripe.comAuthentication is OAuth or Stripe's "agent API keys". Per Stripe's docs, starting in December 2026 it will stop accepting full-access secret keys and restricted keys without the Agent tag; migrate by 30 November 2026. As of November 2026 Stripe MCP is in public preview.
Pitfall: try it against a sandbox first. On a live account, grant only the permissions you need on the OAuth consent screen.
Source: docs.stripe.com/mcp.
aws
AWS Labs' awslabs/mcp repo contains dozens of servers; the general-purpose one is the AWS API server. It runs AWS CLI commands with your credentials.
"aws": {
"command": "uvx",
"args": ["awslabs.aws-api-mcp-server@latest"],
"env": {
"AWS_REGION": "eu-central-1",
"AWS_API_MCP_PROFILE_NAME": "readonly",
"READ_OPERATIONS_ONLY": "true"
}
}Pitfall: READ_OPERATIONS_ONLY is an extra layer; IAM is the real control. Per AWS's own docs, a role with AdministratorAccess permits mutating actions while ReadOnlyAccess allows only reads; use the latter if you only want to inspect.
Source: PyPI awslabs.aws-api-mcp-server 1.6.0 (27 October 2026), github.com/awslabs/mcp.
cloudflare
Instead of one server, Cloudflare runs task-specific remote servers, all Streamable HTTP at /mcp. The old /sse URLs return 410 Gone. Examples: docs https://docs.mcp.cloudflare.com/mcp, Workers observability https://observability.mcp.cloudflare.com/mcp, DNS analytics https://dns-analytics.mcp.cloudflare.com/mcp.
claude mcp add --transport http cloudflare-docs https://docs.mcp.cloudflare.com/mcpThe @cloudflare/mcp-server-cloudflare npm package last shipped in March 2025; don't use it.
Source: github.com/cloudflare/mcp-server-cloudflare (last updated 22 October 2026).
Web and search
brave-search
Web, news, image and local search. Brave's own package replaces the archived one and needs a Brave Search API key. The default transport is stdio.
"brave-search": {
"command": "npx",
"args": ["-y", "@brave/brave-search-mcp-server"],
"env": { "BRAVE_API_KEY": "${BRAVE_API_KEY}" }
}If you prefer Docker: docker run -i --rm -e BRAVE_API_KEY docker.io/mcp/brave-search.
Pitfall: search results are untrusted text; the fetch pitfall below applies here too.
Source: npm @brave/brave-search-mcp-server 2.1.4 (17 September 2026), github.com/brave/brave-search-mcp-server.
fetch
Fetches a URL and returns it as Markdown. A Python package, not on npm.
"fetch": {
"command": "uvx",
"args": ["mcp-server-fetch"]
}Pitfall: two of them. The package's own README warns that the server can reach local and internal IP addresses; factor that in on a company network. And every fetched page is untrusted text entering the model; instructions hidden in a page are a prompt injection path. Don't leave fetch enabled in the same session as servers that can write.
Source: PyPI mcp-server-fetch 2026.11.4 (4 November 2026), github.com/modelcontextprotocol/servers/tree/main/src/fetch.
playwright (browser automation)
Open pages, click, fill forms, take screenshots. Microsoft's Playwright MCP replaces the archived Puppeteer server; it reads the page as structured data through the accessibility tree.
claude mcp add playwright -- npx @playwright/mcp@latest --isolated--isolated keeps the browser profile in memory rather than on disk, so session cookies don't pile up in a persistent profile.
Pitfall: the agent drives a real browser on your behalf, and any text on a page can carry instructions. There are --allowed-origins and --blocked-origins options, but the README states plainly that they are not a security boundary. Don't use it with a profile that is signed in to sensitive accounts; see the Computer Use Agents guide for more.
Source: npm @playwright/mcp 0.0.84 (31 October 2026), github.com/microsoft/playwright-mcp.
AI ecosystem
qdrant (vector database)
Store information in a Qdrant collection and search it semantically; useful in RAG dev loops.
"qdrant": {
"command": "uvx",
"args": ["mcp-server-qdrant"],
"env": {
"QDRANT_URL": "${QDRANT_URL}",
"QDRANT_API_KEY": "${QDRANT_API_KEY}",
"COLLECTION_NAME": "notes"
}
}Pitfall: the default embedding model is sentence-transformers/all-MiniLM-L6-v2. If you filled the collection with a different model, set EMBEDDING_MODEL to match or search results will be meaningless.
Source: PyPI mcp-server-qdrant 0.8.1 (10 December 2025), github.com/qdrant/mcp-server-qdrant. The latest release is about 11 months old; the repo was updated in September 2026.
huggingface
Search models, datasets, Spaces and papers on the Hugging Face Hub. Remote server:
claude mcp add hf-mcp-server -t http "https://huggingface.co/mcp?login"?login signs you in with OAuth; alternatively pass an Authorization: Bearer <HF_TOKEN> header. If you use a token, create a read-only one.
Source: github.com/huggingface/hf-mcp-server, latest release v0.4.33 (4 November 2026).
Systems and devops
docker
Docker's MCP Gateway runs servers from the Docker MCP Toolkit in containers and exposes them to your client as a single endpoint. It is the easy way to run community servers in isolation.
docker mcp gateway run --profile dev-tools
docker mcp client connect claude-code --profile dev-toolsSource: github.com/docker/mcp-gateway (last updated 26 October 2026).
kubernetes
Pods, deployments, logs, Helm. containers/kubernetes-mcp-server uses the context from your kubeconfig. In the current release, runtime settings are passed in a TOML file rather than as command-line flags:
# ~/.config/kubernetes-mcp-server.toml
read_only = true
toolsets = ["core"]"kubernetes": {
"command": "npx",
"args": ["-y", "kubernetes-mcp-server@latest", "--config", "/Users/me/.config/kubernetes-mcp-server.toml"]
}Pitfall: the README itself calls read_only a best-effort guardrail and says Kubernetes RBAC is the access control to rely on. On a production cluster, use a read-only ServiceAccount.
Source: npm kubernetes-mcp-server 0.1.2 (4 November 2026), github.com/containers/kubernetes-mcp-server.
A note on ssh
The previous version of this catalog had an ssh entry. As of November 2026 we couldn't verify an ssh MCP server maintained by a known publisher, so we removed it. Running commands on a remote host is the broadest permission you can grant; if you need it, restrict it to a lab machine and require approval.
Six rules for safe use
- Least privilege: every token, database user, IAM role and OAuth scope is granted at its narrowest. A server's "read-only" flag is an extra layer; the service's own permission system is the real boundary.
- Read-only first: start with GitHub
/readonly, MongoDB--readOnly, AWSREAD_OPERATIONS_ONLY, Kubernetesread_only = trueand a SELECT-only database user. Add write access when you actually need it. - Keep tokens in environment variables: never write a token into a shared
.mcp.json; read it with${VAR}. Prefer OAuth over static tokens where you can. - Known publishers only: install a server from the service's own company or from the
modelcontextprotocolorganization. Check for an npm deprecation notice, GitHub archived status and the latest release date. Being in the Registry or an "awesome" list is not a security review. - Don't trust outside text: web pages, emails, issue comments and search results can contain text that looks like instructions to the model. That is a prompt injection path. Don't keep servers that read outside content enabled in the same session as servers that can write or spend money; require approval for write tools.
- Isolate and switch off: run community servers in Docker or a sandbox. Turn off servers you don't actually use; every enabled server is both attack surface and context usage.
A common starter set
A starter .mcp.json for a software project (Claude Code, project scope). It mixes local and remote servers, and everything starts read-only:
{
"mcpServers": {
"github": {
"type": "http",
"url": "https://api.githubcopilot.com/mcp/readonly",
"headers": { "Authorization": "Bearer ${GITHUB_PAT}" }
},
"sentry": {
"type": "http",
"url": "https://mcp.sentry.dev/mcp"
},
"postgres": {
"command": "npx",
"args": ["-y", "@toolbox-sdk/server", "--prebuilt=postgres", "--stdio"],
"env": {
"POSTGRES_HOST": "${POSTGRES_HOST:-localhost}",
"POSTGRES_DATABASE": "${POSTGRES_DATABASE}",
"POSTGRES_USER": "claude_ro",
"POSTGRES_PASSWORD": "${POSTGRES_PASSWORD}"
}
},
"fetch": {
"command": "uvx",
"args": ["mcp-server-fetch"]
},
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest", "--isolated"]
}
}
}Claude Code's own tools already cover the codebase; these five add the repo host, error tracking, the database, web pages and a browser. Add more when you need them; every server adds tool definitions to the context.
Resources
- MCP Registry (official catalog, preview): registry.modelcontextprotocol.io
- Reference servers: github.com/modelcontextprotocol/servers
- Archived servers: github.com/modelcontextprotocol/servers-archived
- Specification (2026-07-28): modelcontextprotocol.io/specification/2026-07-28
- Claude Code MCP docs: code.claude.com/docs/en/mcp
- Connecting local servers to Claude Desktop: modelcontextprotocol.io/docs/develop/connect-local-servers
Continue reading
- MCP: the protocol itself, and how an agent talks to the outside world.
- Build Your Own MCP Server: set up your own server for a service that isn't on this list.
- Agent Protocols: where MCP sits next to A2A and the agent payment protocols.
- Prompt Injection: why outside content connected through MCP is the biggest risk.
- Claude Skills Guide: MCP's natural companion for daily workflows.
- Function Calling: the underlying tool-use behavior MCP relies on.
- Coding Agents Compared: how each coding agent supports MCP.
Frequently asked questions
Can I still use @modelcontextprotocol/server-github or server-postgres?
Don't. As of November 2026 these packages are deprecated on npm with the notice 'Archived and unmaintained; no security guarantees'. They have been replaced by vendor-maintained servers such as GitHub's own github/github-mcp-server and Google's MCP Toolbox for Databases.
Where does Claude Code store MCP configuration?
Per the Claude Code docs, local- and user-scoped servers live in ~/.claude.json and project-scoped servers live in .mcp.json at the project root. There is no ~/.claude/mcp.json. The recommended way to add a server is the claude mcp add command.
Should I pick a local (stdio) or a remote (HTTP) server?
If the vendor runs a remote server (GitHub, Sentry, Notion, Linear, Stripe and Atlassian do), it is usually easier: nothing to install, you sign in with OAuth, and the vendor ships the updates. Resources on your own machine, such as the filesystem, a local git repo or a local database, need a stdio server.
Is there an official list of MCP servers?
Yes: the MCP Registry at registry.modelcontextprotocol.io. As of November 2026 its repository still describes it as a preview, with the API frozen at v0.1. Being listed is not a security review; you still have to check the publisher yourself.
How do I add a remote MCP server to Claude Desktop?
Remote servers are added as custom connectors: on Pro and Max plans under Customize > Connectors, '+ Add', then 'Add custom connector'; on Team and Enterprise plans an Owner first adds it under Organization settings > Connectors. Local stdio servers go in claude_desktop_config.json.