Cloudflare Python Workers

Cloudflare Python Workers

Status: Beta (requires python_workers compatibility flag) Runtime: Pyodide (Python 3.12+ compiled to WebAssembly) Package Versions: workers-py@1.7.0, workers-runtime-sdk@0.3.1, wrangler@4.58.0 Last Verified: 2026-01-21

Quick Start (5 Minutes)

1. Prerequisites

Ensure you have installed:

• uv - Python package manager
• Node.js - Required for Wrangler

2. Initialize Project

# Create project directory
mkdir my-python-worker && cd my-python-worker

# Initialize Python project
uv init

# Install pywrangler
uv tool install workers-py

# Initialize Worker configuration
uv run pywrangler init

3. Create Entry Point

Create src/entry.py:

from workers import WorkerEntrypoint, Response

class Default(WorkerEntrypoint):
    async def fetch(self, request):
        return Response("Hello from Python Worker!")

4. Configure wrangler.jsonc

{
  "name": "my-python-worker",
  "main": "src/entry.py",
  "compatibility_date": "2025-12-01",
  "compatibility_flags": ["python_workers"]
}

5. Run Locally

uv run pywrangler dev
# Visit http://localhost:8787

6. Deploy

uv run pywrangler deploy

---

Migration from Pre-December 2025 Workers

If you created a Python Worker before December 2025, you were limited to built-in packages. With pywrangler (Dec 2025), you can now deploy with external packages.

Old Approach (no longer needed):

# Limited to built-in packages only
# Could only use httpx, aiohttp, beautifulsoup4, etc.
# Error: "You cannot yet deploy Python Workers that depend on
# packages defined in requirements.txt [code: 10021]"

New Approach (pywrangler):

# pyproject.toml
[project]
dependencies = ["fastapi", "any-pyodide-compatible-package"]

uv tool install workers-py
uv run pywrangler deploy  # Now works!

Historical Timeline:

• April 2024 - Dec 2025: Package deployment completely blocked
• Dec 8, 2025: Pywrangler released, enabling package deployment
• Jan 2026: Open beta with full package support

See: Package deployment issue history

---

Core Concepts

WorkerEntrypoint Class Pattern

As of August 2025, Python Workers use a class-based pattern (not global handlers):

from workers import WorkerEntrypoint, Response

class Default(WorkerEntrypoint):
    async def fetch(self, request):
        # Access bindings via self.env
        value = await self.env.MY_KV.get("key")

        # Parse request
        url = request.url
        method = request.method

        return Response(f"Method: {method}, URL: {url}")

Accessing Bindings

All Cloudflare bindings are accessed via self.env:

class Default(WorkerEntrypoint):
    async def fetch(self, request):
        # D1 Database
        result = await self.env.DB.prepare("SELECT * FROM users").all()

        # KV Storage
        value = await self.env.MY_KV.get("key")
        await self.env.MY_KV.put("key", "value")

        # R2 Object Storage
        obj = await self.env.MY_BUCKET.get("file.txt")

        # Workers AI
        response = await self.env.AI.run("@cf/meta/llama-2-7b-chat-int8", {
            "prompt": "Hello!"
        })

        return Response("OK")

Supported Bindings:

• D1 (SQL database)
• KV (key-value storage)
• R2 (object storage)
• Workers AI
• Vectorize
• Durable Objects
• Queues
• Analytics Engine

See Cloudflare Bindings Documentation for details.

Request/Response Handling

from workers import WorkerEntrypoint, Response
import json

class Default(WorkerEntrypoint):
    async def fetch(self, request):
        # Parse JSON body
        if request.method == "POST":
            body = await request.json()
            return Response(
                json.dumps({"received": body}),
                headers={"Content-Type": "application/json"}
            )

        # Query parameters
        url = URL(request.url)
        name = url.searchParams.get("name", "World")

        return Response(f"Hello, {name}!")

Scheduled Handlers (Cron)

from workers import handler

@handler
async def on_scheduled(event, env, ctx):
    # Run on cron schedule
    print(f"Cron triggered at {event.scheduledTime}")

    # Do work...
    await env.MY_KV.put("last_run", str(event.scheduledTime))

Configure in wrangler.jsonc:

{
  "triggers": {
    "crons": ["*/5 * * * *"]  // Every 5 minutes
  }
}

---

Python Workflows

Python Workflows enable durable, multi-step automation with automatic retries and state persistence.

Why Decorator Pattern?

Python Workflows use the @step.do() decorator pattern because Python does not easily support anonymous callbacks (unlike JavaScript/TypeScript which allows inline arrow functions). This is a fundamental language difference, not a limitation of Cloudflare's implementation.

JavaScript Pattern (doesn't translate):

await step.do("my step", async () => {
  // Inline callback
  return result;
});

Python Pattern (required):

@step.do("my step")
async def my_step():
    # Named function with decorator
    return result

result = await my_step()

Source: Python Workflows Blog

Concurrency with asyncio.gather

Pyodide captures JavaScript promises (thenables) and proxies them as Python awaitables. This enables Promise.all-equivalent behavior using standard Python async patterns:

import asyncio

@step.do("step_a")
async def step_a():
    return "A"

@step.do("step_b")
async def step_b():
    return "B"

# Concurrent execution (like Promise.all)
results = await asyncio.gather(step_a(), step_b())
# results = ["A", "B"]

Why This Works: JavaScript promises from workflow steps are proxied as Python awaitables, allowing standard asyncio concurrency primitives.

Source: Python Workflows Blog

Basic Workflow

from workers import WorkflowEntrypoint, WorkerEntrypoint, Response

class MyWorkflow(WorkflowEntrypoint):
    async def run(self, event, step):
        # Step 1
        @step.do("fetch data")
        async def fetch_data():
            response = await fetch("https://api.example.com/data")
            return await response.json()

        data = await fetch_data()

        # Step 2: Sleep
        await step.sleep("wait", "10 seconds")

        # Step 3: Process
        @step.do("process data")
        async def process_data():
            return {"processed": True, "count": len(data)}

        result = await process_data()
        return result

class Default(WorkerEntrypoint):
    async def fetch(self, request):
        # Create workflow instance
        instance = await self.env.MY_WORKFLOW.create()
        return Response(f"Workflow started: {instance.id}")

DAG Dependencies

Define step dependencies for parallel execution:

class MyWorkflow(WorkflowEntrypoint):
    async def run(self, event, step):
        @step.do("step_a")
        async def step_a():
            return "A done"

        @step.do("step_b")
        async def step_b():
            return "B done"

        # step_c waits for both step_a and step_b
        @step.do("step_c", depends=[step_a, step_b], concurrent=True)
        async def step_c(result_a, result_b):
            return f"C received: {result_a}, {result_b}"

        return await step_c()

Workflow Configuration

{
  "compatibility_flags": ["python_workers", "python_workflows"],
  "compatibility_date": "2025-12-01",
  "workflows": [
    {
      "name": "my-workflow",
      "binding": "MY_WORKFLOW",
      "class_name": "MyWorkflow"
    }
  ]
}

---

Package Management

pyproject.toml Configuration

[project]
name = "my-python-worker"
version = "0.1.0"
requires-python = ">=3.12"
dependencies = [
    "beautifulsoup4",
    "httpx"
]

[dependency-groups]
dev = [
    "workers-py",
    "workers-runtime-sdk"
]

Supported Packages

Python Workers support:

• Pure Python packages from PyPI
• Pyodide packages (pre-built for WebAssembly)

See Pyodide packages list.

HTTP Clients

Only async HTTP libraries work:

# ✅ WORKS - httpx (async)
import httpx

async with httpx.AsyncClient() as client:
    response = await client.get("https://api.example.com")

# ✅ WORKS - aiohttp
import aiohttp

async with aiohttp.ClientSession() as session:
    async with session.get("https://api.example.com") as response:
        data = await response.json()

# ❌ DOES NOT WORK - requests (sync)
import requests  # Will fail!

Requesting New Packages

Request support for new packages at: https://github.com/cloudflare/workerd/discussions/categories/python-packages

---

FFI (Foreign Function Interface)

Access JavaScript APIs from Python via Pyodide's FFI:

JavaScript Globals

from js import fetch, console, Response as JSResponse

class Default(WorkerEntrypoint):
    async def fetch(self, request):
        # Use JavaScript fetch
        response = await fetch("https://api.example.com")
        data = await response.json()

        # Console logging
        console.log("Fetched data:", data)

        # Return JavaScript Response
        return JSResponse.new("Hello!")

Type Conversions

Important: to_py() is a METHOD on JavaScript objects, not a standalone function. Only to_js() is a function.

from js import Object
from pyodide.ffi import to_js

# ❌ WRONG - ImportError!
from pyodide.ffi import to_py
python_data = to_py(js_data)

# ✅ CORRECT - to_py() is a method
async def fetch(self, request):
    data = await request.json()  # Returns JS object
    python_data = data.to_py()   # Convert to Python dict

# Convert Python dict to JavaScript object
python_dict = {"name": "test", "count": 42}
js_object = to_js(python_dict, dict_converter=Object.fromEntries)

# Use in Response
return Response(to_js({"status": "ok"}))

Source: GitHub Issue #3322 (Pyodide maintainer clarification)

---

Known Issues Prevention

This skill prevents 11 documented issues:

Issue #1: Legacy Handler Pattern

Error: TypeError: on_fetch is not defined

Why: Handler pattern changed in August 2025.

# ❌ OLD (deprecated)
@handler
async def on_fetch(request):
    return Response("Hello")

# ✅ NEW (current)
class Default(WorkerEntrypoint):
    async def fetch(self, request):
        return Response("Hello")

Issue #2: Sync HTTP Libraries

Error: RuntimeError: cannot use blocking call in async context

Why: Python Workers run async-only. Sync libraries block the event loop.

# ❌ FAILS
import requests
response = requests.get("https://api.example.com")

# ✅ WORKS
import httpx
async with httpx.AsyncClient() as client:
    response = await client.get("https://api.example.com")

Issue #3: Native/Compiled Packages

Error: ModuleNotFoundError: No module named 'numpy' (or similar)

Why: Only pure Python packages work. Native C extensions are not supported.

Solution: Use Pyodide-compatible alternatives or check Pyodide packages.

Issue #4: Missing Compatibility Flags

Error: Error: Python Workers require the python_workers compatibility flag

Fix: Add to wrangler.jsonc:

{
  "compatibility_flags": ["python_workers"]
}

For Workflows, also add "python_workflows".

Issue #5: I/O Outside Workflow Steps

Error: Workflow state not persisted correctly

Why: All I/O must happen inside @step.do for durability.

# ❌ BAD - fetch outside step
response = await fetch("https://api.example.com")
@step.do("use data")
async def use_data():
    return await response.json()  # response may be stale on retry

# ✅ GOOD - fetch inside step
@step.do("fetch and use")
async def fetch_and_use():
    response = await fetch("https://api.example.com")
    return await response.json()

Issue #6: Type Serialization Errors

Error: TypeError: Object of type X is not JSON serializable

Why: Workflow step return values must be JSON-serializable.

Fix: Convert complex objects before returning:

@step.do("process")
async def process():
    # Convert datetime to string
    return {"timestamp": datetime.now().isoformat()}

Issue #7: Cold Start Performance

Note: Python Workers have higher cold starts than JavaScript. With Wasm memory snapshots (Dec 2025), heavy packages like FastAPI and Pydantic now load in ~1 second (down from ~10 seconds previously), but this is still ~2x slower than JavaScript Workers (~50ms).

Performance Numbers (as of Dec 2025):

• Before snapshots: ~10 seconds for FastAPI/Pydantic
• After snapshots: ~1 second (10x improvement)
• JavaScript equivalent: ~50ms

Mitigation:

• Minimize top-level imports
• Use lazy loading for heavy packages
• Consider JavaScript Workers for latency-critical paths
• Wasm snapshots automatically improve cold starts (no config needed)

Source: Python Workers Redux Blog | InfoQ Coverage

Issue #8: Package Installation Failures

Error: Failed to install package X

Causes:

• Package has native dependencies
• Package not in Pyodide ecosystem
• Network issues during bundling

Fix: Check package compatibility, use alternatives, or request support.

Issue #9: Dev Registry Breaks JS-to-Python RPC

Error: Network connection lost when calling Python Worker from JavaScript Worker Source: GitHub Issue #11438

Why It Happens: Dev registry doesn't properly route RPC calls between separately-run Workers in different terminals.

Prevention:

# ❌ Doesn't work - separate terminals
# Terminal 1: npx wrangler dev (JS worker)
# Terminal 2: npx wrangler dev (Python worker)
# Result: Network connection lost error

# ✅ Works - single wrangler instance
npx wrangler dev -c ts/wrangler.jsonc -c py/wrangler.jsonc

Run both workers in a single wrangler instance to enable proper RPC communication.

Issue #10: HTMLRewriter Memory Limit with Data URLs

Error: TypeError: Parser error: The memory limit has been exceeded Source: GitHub Issue #10814

Why It Happens: Large inline data: URLs (>10MB) in HTML trigger parser memory limits. This is NOT about response size—10MB plain text works fine, but 10MB HTML with embedded data URLs fails. Common with Python Jupyter Notebooks that use inline images for plots.

Prevention:

# ❌ FAILS - HTMLRewriter triggered on notebook HTML with data: URLs
response = await fetch("https://origin.example.com/notebook.html")
return response  # Crashes if HTML contains large data: URLs

# ✅ WORKS - Stream directly or use text/plain
response = await fetch("https://origin.example.com/notebook.html")
headers = {"Content-Type": "text/plain"}  # Bypass parser
return Response(await response.text(), headers=headers)

Workarounds:

• Avoid HTMLRewriter on notebook content (stream directly)
• Pre-process notebooks to extract data URLs to external files
• Use text/plain content-type to bypass parser

Issue #11: PRNG Cannot Be Seeded During Initialization

Error: Deployment fails with user error Source: [Python Workers Redux Blog](https://blog.cloudflare.c

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Kgeminic
• Source: Kgeminic/claude-skills-1
• License: MIT

Install and usage instructions live in the source repository linked above.