Cloudflare MCP Server
★ 3.9k · GitHub →Cloudflare's fleet of remote-hosted MCP servers — Workers, observability, DNS, Radar, and more.
What is Cloudflare MCP Server?
Cloudflare took a different architectural bet from most vendors: instead of one server you install locally, it runs a fleet of domain-specific MCP servers hosted on its own edge at *.mcp.cloudflare.com — documentation, Workers bindings and builds, observability, browser rendering, Logpush, AI Gateway, audit logs, DNS analytics, GraphQL, and several Cloudflare One security servers. Each does one job with a small tool surface.
Because they are remote-hosted, there is nothing to run: you point your MCP client at a URL like https://observability.mcp.cloudflare.com/mcp and authenticate against your Cloudflare account. Clients with native remote-server support (Claude among them) connect directly; older stdio-only clients bridge through the mcp-remote npm package. The older SSE endpoints are deprecated in favor of Streamable HTTP.
It is also a working showcase of what "remote MCP" looks like when done seriously — Cloudflare built much of the tooling for hosting authenticated MCP servers on Workers, then dogfooded it here. The rough edges are real, though: some servers need paid Workers plans, and the observability server can trigger chains of tool calls long enough to blow through a model's context window.
Key capabilities
- ▸ A dozen-plus domain-specific hosted servers — docs, Workers bindings/builds, observability, browser rendering, Logpush, AI Gateway, DNS analytics, audit logs
- ▸ Fully remote: no local process, no version updates — connect to a URL and authenticate with your Cloudflare account
- ▸ Native remote-MCP support in Claude and similar clients; mcp-remote bridges stdio-only clients
- ▸ Narrow per-server tool surfaces keep context cost low when you attach only what you need
- ▸ Built on and validating Cloudflare's own Workers-based remote MCP hosting stack
When to use it
Use these servers if you already run infrastructure on Cloudflare — Workers, DNS, Zero Trust — and want the agent to inspect configs, debug Workers builds, or pull observability data conversationally. Attach only the two or three servers relevant to your stack. If you are not a Cloudflare customer, only the documentation server has any value to you. And for bulk or scripted operations, wrangler and the Cloudflare API remain faster and more predictable than routing everything through an LLM.
Alternatives
Frequently asked questions
How do I connect Claude to the Cloudflare MCP servers?
In clients with native remote-MCP support, add the server URL directly — for example https://docs.mcp.cloudflare.com/mcp or https://bindings.mcp.cloudflare.com/mcp — and complete the Cloudflare login when prompted. For stdio-only clients, configure the mcp-remote npm package as a local bridge that forwards to the hosted URL. Attach servers individually; there is no single URL that exposes everything.
Are the Cloudflare MCP servers free?
The servers themselves are hosted free by Cloudflare and the code is open source. What costs money is the underlying product: features tied to paid plans — some Workers capabilities, Cloudflare One modules — only work if your account has them. The documentation server is fully usable on a free account and needs no special permissions.
Can I run the Cloudflare MCP server locally instead of using the hosted version?
The intended model is remote-hosted — that is the architecture Cloudflare designed around, with OAuth against your account and zero local processes. The source is open, so self-hosting on Workers is possible, but you rarely gain anything: you would maintain deployments and auth yourself to reproduce what the hosted endpoints already do. Local stdio is not the supported path.
Star count updated 2026-07-02 · refreshed weekly from the GitHub API · browse all tools →