Cloudflare API MCP Server

Cloudflare runs a catalog of managed, hosted MCP servers for your own Cloudflare account. The central Cloudflare API MCP Server exposes over 2,500 API endpoints — spanning DNS, Workers, R2, Zero Trust, and more — through just two tools: search() and execute(). Instead of defining a separate tool for every endpoint, the language model writes JavaScript against a typed representation of the OpenAPI spec; the generated code runs inside an isolated Worker sandbox. Per Cloudflare, this "Code Mode" approach uses roughly 1,000 tokens regardless of how many endpoints exist, while a classic MCP server exposing every endpoint as a native tool would consume over a million tokens.

Additional product-specific MCP servers

Alongside the API server, Cloudflare offers around a dozen focused MCP servers for individual products, each reachable under its own subdomain, including: documentation (docs.mcp.cloudflare.com), Workers Bindings (bindings.mcp.cloudflare.com), observability/logs (observability.mcp.cloudflare.com), Radar for Internet traffic data (radar.mcp.cloudflare.com), sandbox containers (containers.mcp.cloudflare.com), browser rendering (browser.mcp.cloudflare.com), AI Gateway (ai-gateway.mcp.cloudflare.com), plus audit logs, DNS analytics, and GraphQL analytics.

Also: the Cloudflare Skills plugin

For Claude Code, Cursor, OpenCode, OpenAI Codex, and other agents that support the Agent Skills standard, the cloudflare/skills plugin bundles the MCP servers together with matching skills and slash commands for working with Cloudflare.

The Code Mode approach in detail

Classic MCP servers define a separate tool with a fixed signature for every action — with over 2,500 Cloudflare API endpoints, that wouldn't be practical within the model's context window. The Code Mode approach flips this: the model receives a typed, searchable representation of the entire API and writes JavaScript code itself, which runs inside an isolated Worker sandbox. For users, this means even very specific, multi-step Cloudflare tasks — such as "update every DNS record in a given zone and log the changes" — can be handled in a single tool call instead of chaining together many individual API calls.

Who the Cloudflare MCP servers fit

The servers are particularly useful for DevOps and infrastructure teams that already actively use Cloudflare for DNS, security, or Workers deployments and want to automate recurring configuration tasks. The product-specific servers also suit focused use cases: the Radar server, for example, for analyzing Internet traffic trends, or the observability server for searching logs during troubleshooting. For a one-off, simple configuration change via the regular Cloudflare dashboard, setting up an MCP client is usually not worth the effort.

Frequently asked questions

Do I need to set up all the product-specific servers at once? No, each server can be connected independently; for most tasks, the central API server or a single focused server is enough. Does access work without OAuth? Yes, for automation an API token is available as an alternative bearer token. Are the servers limited to Zero Trust accounts? No, they work with any Cloudflare account and, depending on permissions, cover DNS, Workers, R2, Zero Trust, and other products.

A Cloudflare account plus an MCP client that supports remote servers over HTTP and OAuth. For CI/CD use, a Cloudflare API token with the required permissions is available as an alternative.

Register the remote server in your MCP client:

{
  "mcpServers": {
    "cloudflare-api": { "url": "https://mcp.cloudflare.com/mcp" }
  }
}

On the first connection, Cloudflare's OAuth dialog opens, where you select the permissions to grant the agent. For CI/CD automation, a Cloudflare API token (user or account token) can be used as a bearer token in the Authorization header instead of OAuth.

For Claude Code, the Skills plugin is an alternative that bundles the MCP servers together with contextual skills:

/plugin marketplace add cloudflare/skills

OAuth authorization for your own Cloudflare account, triggered from the client, where you select the permissions to grant the agent. For automation, a Cloudflare API token (user or account token) can be used as a bearer token instead.

Effective access matches the permissions selected during the OAuth flow or the scopes of the API token used. Via the Code Mode approach, the agent can in principle call any of the 2,500+ Cloudflare API endpoints, limited by those permissions — from read-only analytics access to write changes to DNS, Workers, or security rules.

Requests run through Cloudflare-hosted infrastructure at *.mcp.cloudflare.com. The JavaScript code generated by the language model executes inside an isolated Dynamic Worker sandbox and calls the Cloudflare API from there. Responses contain configuration and analytics data from your own Cloudflare account and are returned to the MCP client and its language model.

Since the server can write to core infrastructure such as DNS, Workers, or Zero Trust rules within the granted permissions, a misdirected agent can cause real outages or security gaps (e.g. faulty DNS changes or loosened firewall rules). Grant only the permissions actually needed during the OAuth flow, use a narrowly scoped API token instead of full account access for automated workflows, and test write actions against production infrastructure in a staging environment first.