Cloudflare opens up 13 remote MCP servers

Cloudflare has made its first batch of remote Model Context Protocol (MCP) servers publicly available. These servers let MCP clients such as Claude.ai, Cursor, Windsurf, and Cloudflare's own AI Playground interact with Cloudflare services through natural language. The initial release covers application development, security, performance, and observability tooling.

The servers are part of a broader push around remote MCP, which Cloudflare and Anthropic are collaborating on to bring remote MCP support to Claude users. Other companies, including Atlassian, PayPal, Sentry, and Webflow, have also built remote MCP servers on Cloudflare's infrastructure.

The 13 servers

Documentation

The Cloudflare Documentation server provides real-time access to up-to-date developer documentation. Instead of relying on potentially stale training data, MCP clients can pull current reference information directly from Cloudflare's docs when answering questions like "Search Cloudflare for the best way to build an AI Agent."

Workers Bindings

The Bindings MCP server gives clients access to application development primitives such as D1 databases, R2 object storage, and Key Value stores. It can read existing resources from an account or create new ones for use in a Workers application. Combined with Cloudflare's base prompt for Workers development, this server is intended to let clients generate full-stack applications from natural language.

Workers Observability

This server integrates with Workers Logs to browse invocation logs and errors, compute statistics across invocations, and find specific invocations matching criteria. It queries logs across all Workers in an account, helping to isolate errors and trends. The telemetry data returned can also feed into new visualizations and broader observability efforts.

Container

The Container MCP server provides MCP clients with a secure, isolated execution environment running on Cloudflare's network. This is useful for clients like claude.ai that lack a built-in development environment: the AI can run its own commands and validate its assumptions in real time while generating code.

Browser Rendering

This server exposes AI-friendly tools from Cloudflare's Browser Rendering REST API for common browser actions, including capturing screenshots, extracting HTML content, and converting web pages to Markdown. These tools are aimed at agents that need to interact with a web browser.

Radar

The Cloudflare Radar MCP server gives clients access to aggregated HTTP traffic data, detail on Autonomous Systems (AS) and IP addresses, traffic anomalies from the Outage Center, trending domains, and domain rank information. It can also generate charts from the data.

Logpush

This server analyzes Logpush job results to give a high-level view of job health. Users can filter and narrow down to specific jobs or scenarios, with queries like "provide me with a list of recently failed jobs" returning human-readable summaries of failures, error messages, and timestamps.

AI Gateway

The AI Gateway MCP server inspects AI Gateway logs and returns details about prompt data and model responses. Example queries can surface average latency for a given account's AI Gateway logs.

AutoRAG

This server makes AutoRAG knowledge bases queryable as MCP tools. Rather than a static one-shot retrieval, agents can dynamically decide when and how often to retrieve information from one or more RAGs, combine it with other tools and APIs, cross-check information, and synthesize a more complete answer. In an example involving blog posts about retrocomputers, an agent queried a specific RAG repeatedly to gather details about restoring an Amiga 1000, searching for specifics on hardware upgrades, DiagROM, and GBA1000 expansions before generating a final guide.

Audit Logs

Audit logs can be large and complex, making it difficult to reconstruct sequences of events. The Audit Logs MCP server allows querying audit logs and generating reports. Typical queries include checking whether anything notable happened in an account around a specific time or whether API keys were used to perform actions. Example: "Were there any suspicious changes made to my Cloudflare account yesterday around lunchtime?"

DNS Analytics

This server provides insight into DNS traffic, helping monitor, analyze, and troubleshoot DNS performance and security across domains. It can review DNS configurations across all domains in an account, access performance reports, and offer recommendations for improvements based on current setup and documentation.

Digital Experience Monitoring

The Digital Experience Monitoring (DEM) MCP server fetches DEM test results to surface performance and availability trends within a Cloudflare One deployment. It provides insights on users, applications, and connected networks, supporting questions about worst user experiences, application performance by time of day, HTTP error patterns, and network path instability.

CASB

The CASB MCP server works with Cloudflare CASB integrations to SaaS and cloud applications. It can explore across users, files, and other asset categories to understand relationships in data that may span many integrations. A common query might be: "Tell me about 'Frank Meszaros' and what SaaS tools they appear to have accessed."

Getting started

Each server is deployed to a URL such as https://observability.mcp.cloudflare.com/sse. Clients with first-class remote MCP support can accept server URLs directly in their interface. With claude.ai, users can navigate to settings, add a new Integration with the server URL, authenticate with Cloudflare, and select the tools the client should be allowed to call.

For clients that do not yet support remote MCP servers, the configuration file (mcp_config.json) needs to be set up using mcp-remote to specify which servers the client can access.

What Cloudflare learned building its own servers

Cloudflare distilled several best practices from building these servers internally, which it shared for other teams building their own:

  • An MCP server is not the whole API schema. The goal is not to wrap the entire API, but to focus on specific jobs and reliable outcomes. A single tool may map to one API or to many. Fewer, more powerful tools can mean smaller context windows, lower cost, faster output, and more valid answers from LLMs.
  • Specialize permissions with multiple servers. Each server operates with precisely the permissions needed for its task, nothing more. Separating capabilities avoids the common security pitfall of over-privileged access.
  • Add robust descriptions within parameters. Detailed tool descriptions help agents understand a parameter's purpose, acceptable values, and impact on server behavior, leading to more intelligent choices instead of arbitrary inputs.
  • Use evals at each iteration. Evaluation tests were implemented for each server to assess whether models followed instructions, selected appropriate tools, and passed correct arguments. This provided a programmatic way to catch regressions when tool descriptions were tweaked.

Users can file issues or requests for new servers in the public GitHub repository. Cloudflare also points builders to a deployable template for creating their own remote MCP server.