The Model Context Protocol turned from an Anthropic specification into an industry standard faster than most people noticed. By the time GitHub launched its MCP Registry at github.com/mcp, the registry already cataloged tens of thousands of community and vendor servers, with the largest ones each at over 100,000 install events and the long tail running into the thousands. (GitHub MCP Registry)
This post is a working tour of the registry, organized by what people actually do for a living. Each section names the servers I would reach for first in that domain, says what they unlock, cites the official source, and shows a copy-paste configuration block you can drop into Claude Desktop or Claude Code. I have tried to keep this post free of the usual MCP hype. The point is to ship work faster, not to collect logos.
A 30-second MCP recap
An MCP server is a small program that exposes tools, resources, and prompts to an MCP client (Claude Code, Claude Desktop, Cursor, and a growing list). Once a client connects to a server, the model can call those tools like functions inside a conversation. Reading a Notion page, querying Postgres, opening a Linear ticket, scraping a webpage, all become calls the model can make on its own. The protocol itself is documented at modelcontextprotocol.io and the official organization at github.com/modelcontextprotocol. (MCP organization on GitHub, Example servers)
There are three places to discover servers, in roughly increasing strictness:
- GitHub MCP Registry. The newest hub. Counts installs, tags by category, and includes both vendor and community servers.
- modelcontextprotocol/servers for the official reference implementations: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, and Time. These are educational templates more than production endpoints.
- Vendor-maintained repositories. The big ones (GitHub, Microsoft, Atlassian, Stripe, Sentry, Cloudflare, Notion) ship their own. These are the production-grade servers most teams should pick first.
A practical note before the list: where a vendor publishes its own server, prefer it. The vendor server is the one that gets updated when the underlying API changes, the one that handles auth properly, and the one that does not expose the model to a third party reading your tokens. Community servers are useful when no vendor server exists or when the vendor's surface is too narrow.
Software engineering and DevOps
This is the densest category in the registry, because developers built MCP for themselves first.
github/github-mcp-server. GitHub's official server. Repos, issues, pull requests, code search, secret scanning, security advisories. The single highest-leverage MCP install for any code-adjacent workflow. Gives Claude the ability to triage issues, open targeted PRs, and read across an organization's repos with the same auth a person has.
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
}
}
}
microsoft/azure-devops-mcp. Azure Boards, Repos, Pipelines, Test Plans. The right server for shops on the Microsoft side of the dev tooling fence. (microsoft/mcp catalog)
microsoft/playwright-mcp. Playwright via accessibility snapshots. Lets the model drive a real browser, take page snapshots, fill forms, and validate flows. The cleanest way to give a coding agent end-to-end testing capability without wiring a custom harness.
microsoft/markitdown. Convert PDFs, Word, Excel, PowerPoint, audio transcripts, and images to clean Markdown. Reported north of 121,000 installs on the registry. The first server to install on any docs-heavy workstation. (Microsoft Markitdown)
Sentry via getsentry/sentry-mcp. Query issues, events, releases, projects from inside an IDE or assistant. For a team that spends Mondays triaging the weekend's alerts, this server is a substantial weekly time save. (Sentry MCP server)
Cloudflare Workers, KV, R2, D1, Hyperdrive. Cloudflare's own MCP coverage was launched alongside its agents-SDK push, and is the right way for any team building on Workers to give its assistant access to platform state. (Cloudflare MCP demo day)
For prototypes, the official Filesystem, Git, and Fetch reference servers are worth attaching to any new project. They are deliberately simple, but they remove the most common "Claude can't see this file" failure mode in a fresh repo. (modelcontextprotocol/servers reference set)
Project management and team operations
Linear. Linear shipped its own MCP server in 2025, with an intentionally tight surface: search issues, create issues, update status, comment. Linear's approach is interesting in that they treat MCP as the agent surface for their own product and have publicly bet on the agentic-PM direction. (Linear MCP docs)
Atlassian (Jira and Confluence Cloud). Atlassian ships a remote MCP server hosted on Cloudflare. Jira issue creation and search, Confluence page reading and writing, with full OAuth. The right place to start for teams already paying Atlassian. (Atlassian announcement)
Notion. MCP connections for Custom Agents. Read, search, and update Notion pages and databases. Notion's surface is genuinely useful as a documentation system that the model can keep current. (Notion MCP help)
Slack. Slack's own MCP server lets agents search channels, post messages, fetch threads, and resolve users. Use the official server, not the older community one, for production work. (Slack MCP docs)
For a small team, the Linear plus GitHub plus Slack triple covers most of "what is the team working on" without further configuration.
Customer support and CRM
This is a category where MCP is changing how reps work, not just how engineers work.
Salesforce. The Salesforce MCP server exposes core CRM operations: read accounts, contacts, opportunities, create and update records. Useful even at small scale because most Salesforce dashboards are slow to assemble manually. (Salesforce MCP coverage in skyvia 2026 list)
HubSpot. Similar surface to Salesforce, on the HubSpot side. Contacts, deals, marketing campaigns. The HubSpot MCP server is one of the most-installed in the registry's CRM category as of late 2025.
Intercom. One of the original MCP launch partners with Cloudflare. Lets the model search conversations, draft responses, and resolve threads. (Cloudflare MCP demo day for launch partners)
Zendesk. A community MCP server has existed for some time; in late 2025 Zendesk announced its own. Search tickets, post replies, escalate.
For a 30-person company, a useful pattern is to expose the support and CRM servers to a single internal "ops" agent that can summarize the morning's tickets, surface CRM mismatches, and post a daily digest to Slack. None of those are LLM-native; all of them become trivial once the agent can talk to the underlying systems.
Finance and payments
This is a small but high-value category, and it is where the trader-adjacent reader of the Quadrature and Radix post on this site will find the most utility.
Stripe. Stripe's official MCP server, launched in the Cloudflare cohort, exposes customers, charges, subscriptions, and payouts. For founders and ops teams, this is the cleanest way to expose Stripe state to a Claude conversation without reinventing the dashboard. (Cloudflare launch coverage)
Block. Block (formerly Square) published its own MCP server alongside Stripe's launch. Different surface, same idea.
PayPal. PayPal also shipped during the Cloudflare cohort. Payment search, dispute lookup, payout creation.
Quant and trading data. Vendor-side coverage is still thin. The official Coinbase Advanced API is wrapped by community MCP servers but does not have an Anthropic- or Coinbase-blessed first-party MCP yet. For now, the right pattern for trading work is to expose your own internal trading tools as an MCP server using the official Python SDK. The companion post The Retail Quant Stack shows the pieces; wrapping them as an MCP server adds about 60 lines of code via modelcontextprotocol/python-sdk.
# Minimal MCP server exposing a backtest runner
from mcp.server.fastmcp import FastMCP
from tsmom_backtest import run as run_tsmom_backtest
mcp = FastMCP("retail-quant")
@mcp.tool()
def run_backtest() -> dict:
"""Run the time-series momentum backtest and return summary stats."""
return run_tsmom_backtest()
if __name__ == "__main__":
mcp.run()
Data, databases, and analytics
PostgreSQL. The official Postgres MCP server in the modelcontextprotocol reference set provides read-only schema inspection and parameterized queries. Lock it to a read-only role in production. (servers-archived/postgres)
Supabase. Supabase shipped its own MCP server, and it is materially better than running the raw Postgres server because it understands Supabase's row-level security model.
Microsoft Fabric, SQL Server, Dataverse. Microsoft's enterprise data line is broad and growing, with each major product carrying its own MCP. Useful primarily for shops that have already standardized on the Microsoft data stack. (microsoft/mcp)
Snowflake, BigQuery, Databricks. All three have community MCP servers; vendor-blessed servers are in active development as of mid-2026. Treat the community servers as research tools, not production data conduits, until the vendor ones land.
For analytics, Microsoft Clarity MCP is one of the underrated wins. It lets a model query session-level analytics directly. (clarity-mcp-server)
Marketing, content, and SEO
Ahrefs MCP. Domain Rating, backlink profile, keyword research, competitive intelligence. Listed in the 2026 top servers list and the most useful single SEO server.
Brave Search. The official Brave Search MCP server is in the modelcontextprotocol reference set. Use it as the lightweight default for any "go look this up" tool call. (Brave Search reference server)
Firecrawl. Web data extraction at scale. The right tool when you need to crawl a site rather than fetch a single page. (Firecrawl on GitHub MCP Registry)
Brightdata. Web search and navigation for AI agents. Useful when you need anti-bot-resilient data acquisition. The category is sensitive; understand the legal and ToS posture of any target before scraping it.
For small marketing teams: Brave for the morning brief, Ahrefs for the weekly competitive scan, and Notion for storing what you learn. That stack runs free or near-free on personal plans and replaces a meaningful slice of paid tooling.
Compliance, security, and observability
Sentry MCP. Already mentioned above; the same server doubles as the operations team's incident-postmortem assistant.
Microsoft Sentinel Data Exploration MCP. Lets a security analyst query SIEM data conversationally. (microsoft/mcp)
Netdata. Real-time infrastructure monitoring with metrics, logs, alerts, and ML-based anomaly detection. Reported to be one of the highest-installed MCP servers on the GitHub Registry, with 78,000 plus installs as of late 2025. (GitHub MCP Registry)
Endpoint security audit. This is one I built and maintain myself for the Trellix-style endpoint posture audits that Windows admins ask about. It is a thin wrapper around a few PowerShell calls and the right shape if your fleet is on Trellix or another major endpoint stack and you want the model to reason about an audit log without you copy-pasting one.
Industry-specific niches worth knowing about
Healthcare. No vendor-blessed MCP server has shipped for the major EHR vendors as of mid-2026. Community servers exist around FHIR proxies; treat them as research-grade only and never connect them to a system carrying PHI without an enterprise security review.
Legal. Westlaw, Lexis, Bloomberg Law have not published MCP servers. The serious option for legal-tech teams is to wrap a Practice Management API (Clio, MyCase) yourself; both have well-documented REST surfaces and the MCP wrapping is straightforward.
Real estate. Zillow, Redfin, MLS data are gated by ToS that materially limits what an MCP server can lawfully expose. Where vendors have sanctioned APIs (Zillow Premier Agent, ATTOM Data, RealtyMole) the wrapping is again straightforward, but check the ToS before exposing any scrape-derived data.
Manufacturing and field operations. ServiceNow, SAP, and IBM have all shipped or announced MCP coverage. (IBM MCP catalog, Oracle MCP)
Configuration patterns that save time
If you are putting a stack of these together, three patterns are worth respecting.
1. Per-project .mcp.json instead of a single global config. Each project gets only the servers it actually needs. The model is faster, the auth surface is smaller, and the per-project secret is less likely to leak.
2. Read-only credentials by default. When a server can be read-only (Postgres, Stripe, GitHub, Linear), use a read-only token. Upgrade to write only for the specific projects where the agent is supposed to write.
3. The Cloudflare-hosted remote MCP path for vendors that support it. Atlassian, Stripe, Sentry, and several others run Cloudflare-hosted remote MCP endpoints. These avoid the local-process management headache and handle OAuth properly. Use the remote path if you can. (Cloudflare MCP launch coverage)
A representative claude_desktop_config.json for an engineering-and-ops user looks like this:
{
"mcpServers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
},
"postgres": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-postgres",
"postgresql://readonly_user:pwd@localhost/mydb"]
},
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem",
"/Users/me/projects"]
},
"fetch": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-fetch"]
},
"sentry": {
"command": "npx",
"args": ["-y", "@sentry/mcp-server"],
"env": { "SENTRY_AUTH_TOKEN": "..." }
}
}
}
Tighten or relax to match your stack. The five above will move work for most engineers immediately.
What is genuinely missing as of mid-2026
A clear-eyed list, because it informs which MCP servers are worth waiting for.
- Native, vendor-blessed crypto exchange MCPs (Coinbase, Binance, Kraken). Community wrappers exist; the official ones are not yet shipped at the time of writing.
- Bloomberg Terminal MCP. Most desk-side users would buy this in a heartbeat.
- EHR vendors in healthcare, for the regulatory reasons noted above.
- Major LLM-evaluation platforms. A vendor MCP for Braintrust, LangSmith, or Helicone would close a real gap in evaluation workflows.
- Standardized observability across vendors. Right now, each observability vendor ships its own server. A neutral OpenTelemetry-based MCP would let teams switch backends without rewriting the agent layer.
If any of these land in 2026 they will move work materially. If not, the right move is to wrap your own internal API as a small MCP server using the Python SDK or TypeScript SDK. A useful internal MCP server is usually under 200 lines of code.
Related reading
- Inside Quadrature Capital and Radix Trading for the trading-firm context
- The Retail Quant Stack for the runnable backtest you can wrap as an MCP server
- Why Top Quant Desks Outperform And How U.S. Tax Code Multiplies The Edge for the structural side of the trading economy
- What Actually Fixed My Claude Code Sessions for the broader prompt and process discipline that benefits from MCP integration
Fact-check notes and sources
- GitHub MCP Registry: github.com/mcp
- Official MCP organization, reference servers, and SDKs: github.com/modelcontextprotocol, modelcontextprotocol/servers, modelcontextprotocol/python-sdk, example servers
- GitHub MCP Server: github/github-mcp-server
- Microsoft MCP catalog: microsoft/mcp, microsoft/azure-devops-mcp, microsoft/playwright-mcp, microsoft/markitdown, microsoft/clarity-mcp-server
- IBM and Oracle MCP catalogs: IBM/mcp, oracle/mcp
- Sentry MCP: getsentry/sentry-mcp
- Atlassian remote MCP: Inside Atlassian announcement
- Cloudflare MCP launch and partners: Cloudflare MCP demo day
- Linear MCP: Linear docs, The Register coverage
- Notion MCP: Notion help center
- Slack MCP: Slack developer docs
- Curated server lists: habitoai/awesome-mcp-servers, skyvia top MCP servers 2026, Premai 25 best MCP servers, Merge.dev examples
This post is informational. MCP servers grant a model real access to systems with real data, real cost, and real liability surface. Use vendor-published servers where available, scope tokens to read-only by default, store secrets outside committed config, and never expose a write-capable server to a public-facing agent without a security review. Past install counts on the registry are not a guarantee of code quality. Consult your security team before connecting any MCP server to systems carrying customer data.
