Cloudflare

Unified skill for building on the Cloudflare platform. Covers deployment, sandboxed code execution, and MCP client integration in Workers.

Prerequisites

• When sandboxing blocks deployment network calls, rerun with sandbox_permissions=require_escalated.
• Deployments may require extended timeout values.

Authentication (Required Before Deploy)

Verify auth before wrangler deploy, wrangler pages deploy, or npm run deploy:

npx wrangler whoami    # Shows account if authenticated

Not authenticated? See references/wrangler/auth.md

• Interactive/local: wrangler login (one-time OAuth)
• CI/CD: Set CLOUDFLARE_API_TOKEN env var
• Sandbox SDK local dev: Set CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID in dev.vars

Part 1: Deploy (Workers, Pages, and Platform Services)

Use the decision trees below to find the right product, then load detailed references.

Quick Decision Trees

"I need to run code"

Need to run code?
├─ Serverless functions at the edge → workers/
├─ Full-stack web app with Git deploys → pages/
├─ Stateful coordination/real-time → durable-objects/
├─ Long-running multi-step jobs → workflows/
├─ Run containers → containers/
├─ Multi-tenant (customers deploy code) → workers-for-platforms/
├─ Scheduled tasks (cron) → cron-triggers/
├─ Lightweight edge logic (modify HTTP) → snippets/
├─ Process Worker execution events (logs/observability) → tail-workers/
└─ Optimize latency to backend infrastructure → smart-placement/

"I need to store data"

Need storage?
├─ Key-value (config, sessions, cache) → kv/
├─ Relational SQL → d1/ (SQLite) or hyperdrive/ (existing Postgres/MySQL)
├─ Object/file storage (S3-compatible) → r2/
├─ Message queue (async processing) → queues/
├─ Vector embeddings (AI/semantic search) → vectorize/
├─ Strongly-consistent per-entity state → durable-objects/ (DO storage)
├─ Secrets management → secrets-store/
├─ Streaming ETL to R2 → pipelines/
└─ Persistent cache (long-term retention) → cache-reserve/

"I need AI/ML"

Need AI?
├─ Run inference (LLMs, embeddings, images) → workers-ai/
├─ Vector database for RAG/search → vectorize/
├─ Build stateful AI agents → agents-sdk/
├─ Gateway for any AI provider (caching, routing) → ai-gateway/
└─ AI-powered search widget → ai-search/

"I need networking/connectivity"

Need networking?
├─ Expose local service to internet → tunnel/
├─ TCP/UDP proxy (non-HTTP) → spectrum/
├─ WebRTC TURN server → turn/
├─ Private network connectivity → network-interconnect/
├─ Optimize routing → argo-smart-routing/
├─ Optimize latency to backend (not user) → smart-placement/
└─ Real-time video/audio → realtimekit/ or realtime-sfu/

"I need security"

Need security?
├─ Web Application Firewall → waf/
├─ DDoS protection → ddos/
├─ Bot detection/management → bot-management/
├─ API protection → api-shield/
├─ CAPTCHA alternative → turnstile/
└─ Credential leak detection → waf/ (managed ruleset)

"I need media/content"

Need media?
├─ Image optimization/transformation → images/
├─ Video streaming/encoding → stream/
├─ Browser automation/screenshots → browser-rendering/
└─ Third-party script management → zaraz/

"I need infrastructure-as-code"

Need IaC? → pulumi/ (Pulumi), terraform/ (Terraform), or api/ (REST API)

Product Index

Compute & Runtime

Storage & Data

AI & Machine Learning

Networking & Connectivity

Security

Media & Content

Real-Time Communication

Developer Tools

Infrastructure as Code

Other Services

Part 2: Sandbox SDK (Secure Code Execution)

Build secure, sandboxed applications that execute untrusted code safely using Cloudflare's Sandbox SDK.

Official docs:

• https://sandbox.cloudflare.com
• https://developers.cloudflare.com/sandbox/

When to Use

• Run untrusted code safely (CI/CD tasks, code review, evals, custom agents)
• Provide a code interpreter / REPL-like experience
• Spin up ephemeral environments (Node/Python/other images) to run scripts
• Proxy traffic to an app running inside a sandbox (preview URLs)

Setup

1. Install the SDK:

npm i @cloudflare/sandbox

2. Add the Sandbox service binding to wrangler.jsonc:

{
  "services": [
    { "binding": "SANDBOX", "service": "sandbox" }
  ]
}

3. Set credentials for local dev in dev.vars:

CLOUDFLARE_API_TOKEN=...
CLOUDFLARE_ACCOUNT_ID=...

Core Workflow

Get (or create) a sandbox

import { getSandbox } from '@cloudflare/sandbox';

const sandbox = await getSandbox(env.SANDBOX, {
  id: 'stable-session-id',  // Reuse across requests (optional)
  image: 'nodejs',           // Environment (optional)
  memoryMB: 256,             // Limits (optional)
  timeoutMs: 60_000,
});

To reuse the same sandbox across requests, pass a stable id (e.g. per user/session).

Run a command

const result = await sandbox.exec('echo hello world');
// result: { exitCode, stdout, stderr }

Write files and execute scripts

await sandbox.mkdir('/tmp/app');
await sandbox.writeFile('/tmp/app/index.js', 'console.log("hi from sandbox")');
await sandbox.exec('node /tmp/app/index.js');

Stream output for long-running tasks

Use execStream(...) when you want progressive output (SSE):

const sse = await sandbox.execStream('bash -lc "for i in 1 2 3; do echo $i; sleep 1; done"');
return new Response(sse, { headers: { 'content-type': 'text/event-stream' } });

Manage background processes

const { pid } = await sandbox.startProcess('node server.js');
const procs = await sandbox.listProcesses();
await sandbox.killProcess(pid);

Expose ports / proxy requests

• Use listenPort(port) to obtain a preview URL.
• Use proxyToSandbox(request, sandbox, { port, pathPrefix }) to forward Worker traffic to the sandbox.

Security Checklist

• Keep network disabled unless you explicitly need it (network: true per exec/process).
• Set strict resource limits (CPU/memory/timeouts) appropriate to your use case.
• Avoid interpolating user input into shell commands; prefer templates/allowlists.
• Treat /tmp as ephemeral. Use mount(...) for persistence.

Sandbox SDK References

• references/sandbox-sdk/api-quick-ref.md -- API quick reference
• references/sandbox-sdk/examples.md -- Code examples
• references/sandbox/ -- Platform-level sandbox reference (config, gotchas, patterns)

Sandbox SDK Troubleshooting

• "SANDBOX is undefined": missing services binding in wrangler.jsonc.
• 401 / auth failures: missing/invalid CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID.
• No output / truncated output: prefer execStream(...) for large logs.

Part 3: MCP Client in Workers (Agents SDK)

Build AI agents in Cloudflare Workers that connect to remote MCP (Model Context Protocol) servers for tool access.

Official docs:

• https://developers.cloudflare.com/agents/model-context-protocol/
• https://developers.cloudflare.com/agents/guides/connect-mcp-client/
• AI SDK v6 Migration Guide: https://ai-sdk.dev/docs/migration-guides/migration-guide-6-0

When to Use

• Build a Cloudflare Worker that acts as an MCP client
• Connect an AI agent to remote MCP servers (Streamable HTTP transport)
• Use the Vercel AI SDK with Cloudflare's Agents framework
• Implement tool-calling agents with external MCP tool providers

Step 1: Use the Correct Package

The @cloudflare/agents package is deprecated. Use the community agents package:

npm install agents@^0.3.6

// CORRECT
import { Agent, getAgentByName, routeAgentRequest } from 'agents';

// DEPRECATED - DO NOT USE
// import { Agent } from '@cloudflare/agents';

Step 2: AI SDK v6 Breaking Changes

AI SDK v6 replaced maxSteps with stopWhen and a step counter function:

// BEFORE (AI SDK v5)
const result = await generateText({
  model: anthropic('claude-sonnet-4@latest'),
  tools: mcpTools,
  maxSteps: 10,
  prompt: userMessage,
});

// AFTER (AI SDK v6)
import { generateText, stepCountIs } from 'ai';

const result = await generateText({
  model: anthropic('claude-sonnet-4@latest'),
  tools: mcpTools,
  stopWhen: stepCountIs(10),  // Default is now 20 steps
  prompt: userMessage,
});

Step 3: MCP Client Manager API

The Agent class has an mcp property with MCPClientManager. Key details:

• Use this.mcp.mcpConnections (not mcpServers)
• MCPClientConnection has url (URL object) and connectionState properties
• Use this.mcp.getAITools() to get tools for AI SDK

export class MyAgent extends Agent<Env, AgentState> {
  getMcpStatus() {
    const connections = this.mcp.mcpConnections;

    for (const [id, connection] of Object.entries(connections)) {
      if (connection.url.toString().includes("my-server")) {
        return {
          connected: connection.connectionState === "ready",
          serverId: id,
          state: connection.connectionState,
        };
      }
    }
    return { connected: false };
  }

  async callWithTools(prompt: string) {
    const tools = this.mcp.getAITools();
    // Use tools with generateText...
  }
}

Step 4: Fix AgentNamespace Type Issues

The AgentNamespace generic doesn't properly infer agent types with getAgentByName. Use type assertions as a workaround:

import { getAgentByName } from 'agents';

const agent = (await getAgentByName(
  env.MY_AGENT as any,
  "agent-name"
)) as unknown as MyAgent;

const result = await agent.myCustomMethod();

Step 5: Workers Compatibility Date

Set compatibility date to 2024-09-23 or later for automatic node: prefix resolution:

# wrangler.toml
name = "my-agent"
compatibility_date = "2024-09-23"
compatibility_flags = ["nodejs_compat"]

[[durable_objects.bindings]]
name = "MY_AGENT"
class_name = "MyAgent"

[[migrations]]
tag = "v1"
new_sqlite_classes = ["MyAgent"]

Without this, imports like path from dependencies (e.g., mime-types) will fail with:

Could not resolve "path"
The package "path" wasn't found on the file system but is built into node.

MCP Client Verification

After implementation:

1. Build succeeds: wrangler deploy --dry-run
2. Local dev works: wrangler dev --local
3. MCP connection: Check logs for successful connection to MCP server
4. Tool discovery: Verify this.mcp.getAITools() returns expected tools

wrangler dev --local --port 8799
curl http://localhost:8799/
curl http://localhost:8799/api/status

MCP Client Notes

• Package confusion: @cloudflare/agents was an early experiment; agents is actively maintained
• AI SDK migration: v6 was a major refactor; check migration guide for other breaking changes
• MCP OAuth: Use this.addMcpServer(name, url, callbackHost) for OAuth-protected MCP servers -- returns authUrl if authentication needed
• Connection persistence: MCP connections are persisted in Agent's SQL storage and automatically restored on restart

Troubleshooting

Escalated Network Access

If deployment fails due to network issues (timeouts, DNS errors, connection resets), rerun the deploy with escalated permissions (use sandbox_permissions=require_escalated).

Common Issues