The registry just crossed a meaningful threshold
As of August 6, 2026, the Model Context Protocol registry indexes nearly 20,000 MCP servers. That number matters less as a milestone than as a signal: MCP is no longer a niche developer experiment. It is the de facto standard for connecting AI assistants to external tools, and Claude Cowork’s plugin system is built around it.
The practical implication is that the integration gap between “what your official plugins cover” and “every tool your team actually uses” is now closeable for almost any stack — including the internal dashboards and niche CRMs that no vendor will ever build a native connector for.
What MCP is — and why it replaced one-off integrations
Before MCP, every integration between an AI assistant and an external app required custom, one-off wiring. A connector for Slack looked nothing like a connector for Notion. Maintaining them was a tax on every team that built them.
MCP standardises that plumbing. An MCP server is a small program that exposes tools, resources, and prompt templates to any MCP-compatible client — including Claude Cowork. One protocol, any app. Which is why, when a new service launches, an MCP server typically appears within days rather than months.
In practice, MCP servers are how you give Claude Cowork new capabilities: read your calendar, query a database, search your company wiki, or control a browser. Anthropic does not need to build each integration. The community — and your own engineering team — can.
Who this affects most directly
Three groups feel this most directly.
Knowledge workers on Pro or Max plans gain access to a long tail of apps their official plugins don’t yet cover. MCP server support is not plan-gated — any Pro or Max subscriber can connect servers. The plan determines usage limits, not access.
Teams with internal tools can expose proprietary APIs, databases, and dashboards as MCP servers in an afternoon. That turns Claude Cowork from a general assistant into something that understands your specific stack.
Operators running Claude Code will find that most MCP servers built for Claude Code work with Cowork too, because both share underlying technology. The config block — the mcpServers entry — ports over almost verbatim.
What the official connectors already cover
Before reaching for a community server, check whether an official connector already exists. Official connectors are maintained by the vendor, documented, and receive regular updates — they are the safest starting point.
Built into Claude Cowork via Settings → Integrations: Google Drive, Gmail, GitHub, Excel, PowerPoint, and the local file system. The Productivity plugin extends that list to Slack, Notion, Asana, Linear, Jira, Monday, ClickUp, and Microsoft 365. The Enterprise Search plugin adds Guru. If your tool appears on that list, you don’t need a community server — the plugin already handles it.
Community servers become relevant for everything else: niche CRMs, internal dashboards, design tools, and any service that launched recently enough that no official connector exists yet.
The community server categories worth your attention
The Claude Cowork MCP directory tracks the official registry and sorts by stars. Based on consistent download patterns, four categories stand out.
Databases — Postgres, MySQL, and SQLite servers with read-only defaults. This is the single highest-leverage category for most teams: point a server at your analytics database and ask for a formatted weekly report, formulas included.
File system and documents — structured access to folders, PDFs, spreadsheets, and markdown vaults. These are the most downloaded servers in the registry for a reason.
Web browsing and search — fetch URLs, read pages, and run searches with proper rate-limit handling. Useful for research workflows where Cowork needs to synthesise sources it can’t already see.
Design tools — Figma, Draw.io, and image-generation servers. These turn Cowork into a design assistant that can read a Figma file, generate the export, and drop it into a repository.
When picking any community server: check GitHub stars as a popularity signal (not a safety guarantee), read the repository before installing, prefer servers with active recent commits, and avoid anything requesting permissions broader than the task requires.
How to install a server in five steps
The installation flow is consistent across most servers.
- Find the server — browse the MCP directory or the official registry. Copy the package name (npm, pip, or uv).
- Install it locally — run
npm install -g <package>oruvx <package>. Remote servers need an endpoint URL instead of a local install. - Add it to Cowork — open Settings → Integrations → MCP Servers, click Add, and paste the launch command or URL.
- Grant minimal permissions — Cowork will ask which folders or resources the server may access. Start narrow and expand only when a task requires it.
- Test with a simple prompt — “List my tasks in [app]” — before moving to real workflows.
The security trade-off you need to accept consciously
The strongest objection to broad MCP adoption is legitimate: community servers vary widely in quality and maintenance, and a poorly scoped server is a real attack surface. That concern does not go away because the ecosystem is large — if anything, a larger registry increases the variance.
MCP servers get real access to real systems. Remote HTTP/SSE servers send your data to their host. Only use ones you trust. Local servers are safer by default, but still warrant a read of the source code before installation. Treat every community MCP server the way you’d treat any third-party dependency: audit it, pin a version, and keep it updated.
The Claude Cowork security and privacy guide covers sandboxing and permission best practices in more depth. Read it before connecting anything to a production database.
One action this week
Open the MCP directory, filter by stars, and install one server before Friday. If your team runs any analytics work, start with the database category: a read-only Postgres or SQLite server is the lowest-risk, highest-return entry point, and the setup takes roughly 20 minutes. Once that’s working, layer in the official plugins for the tools already on the supported list — and Claude Cowork shifts from a standalone assistant to the control centre of your workflow.
— Eagentix
Eagentix helps growth-focused enterprises redesign and automate manual business processes. We combine executive strategy, implementation support, and managed services to build dependable operations across Southeast Asia.