Devkit Anthropic Proxy

A CloudFlare Worker that proxies multiple API formats to A4F's API gateway, enabling tools like Roo Code, Cline, and other AI clients to work seamlessly with A4F's backend.

Features

Anthropic API (Messages)

✅ POST /v1/messages - Main Anthropic Messages API endpoint
✅ POST /v1/messages/count_tokens - Token counting endpoint
✅ Full streaming support with proper SSE event conversion
✅ Handles system prompts, multi-modal content (images), and tool calling
✅ Near accurate Claude token counting via @lenml/tokenizer-claude

OpenAI API (Chat Completions)

✅ POST /v1/chat/completions - OpenAI Chat Completions API (pass-through)
✅ Full streaming support (SSE pass-through)
✅ Direct forwarding to A4F backend

OpenAI Responses API

✅ POST /v1/responses - OpenAI Responses API for GPT-5.1 Codex models
✅ Streaming and non-streaming support
✅ Automatic model prefix handling

Common Features

✅ GET /v1/models - List available models (Claude + GPT Codex only)
✅ GET /health - Health check endpoint
✅ Dual authentication: x-api-key header (Anthropic) or Authorization: Bearer (OpenAI)
✅ CORS support for browser-based clients
✅ Path normalization (handles /v1/v1/ and // prefixes)
✅ Flexible path matching (supports both /v1/* and /* paths)

API Endpoints

Authentication

All endpoints (except /health) require authentication via one of:

x-api-key: YOUR_API_KEY header (Anthropic style)
Authorization: Bearer YOUR_API_KEY header (OpenAI style)

Model Name Handling

Different endpoints handle model names differently:

Endpoint	Client Sends	Proxy Adds Prefix	Example
`/v1/messages`	Model name only	`provider-7/`	`claude-sonnet-4-20250514` → `provider-7/claude-sonnet-4-20250514`
`/v1/chat/completions`	Full model ID	None (pass-through)	`provider-7/claude-sonnet-4-20250514`
`/v1/responses`	Model name only	`provider-5/`	`gpt-5.1-codex` → `provider-5/gpt-5.1-codex`
`/v1/models`	N/A	Strips prefixes	Returns `claude-sonnet-4-20250514`, `gpt-5.1-codex`

Health Check

curl http://localhost:8787/health

Response:

{"status":"ok","service":"devkit-anthropic-proxy"}

List Models

Returns available Claude models (from provider-7) and GPT Codex models (from provider-5) with prefixes stripped.

Note: Only GPT Codex models are included (not all GPT models) because they support the /v1/responses API. Non-codex GPT models like gpt-4o only support /v1/chat/completions and should be accessed directly via that endpoint.

curl http://localhost:8787/v1/models \
  -H "Authorization: Bearer YOUR_API_KEY"

Response:

{
  "object": "list",
  "data": [
    {"id": "claude-sonnet-4-20250514", "object": "model", "owned_by": "anthropic"},
    {"id": "claude-opus-4-5-20251101", "object": "model", "owned_by": "anthropic"},
    {"id": "gpt-5.1-codex", "object": "model", "owned_by": "openai"},
    {"id": "gpt-5.1-codex-mini", "object": "model", "owned_by": "openai"}
  ]
}

Anthropic Messages API (`/v1/messages`)

Converts Anthropic API format to OpenAI format, forwards to A4F, and converts the response back.

Non-Streaming:

curl -X POST http://localhost:8787/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "max_tokens": 100,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Streaming:

curl -X POST http://localhost:8787/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "max_tokens": 100,
    "stream": true,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

OpenAI Chat Completions API (`/v1/chat/completions`)

Pass-through endpoint for OpenAI-format requests. Model names are NOT modified - you must include the full provider prefix.

Non-Streaming:

curl -X POST http://localhost:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "provider-7/claude-sonnet-4-20250514",
    "max_tokens": 100,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Streaming:

curl -X POST http://localhost:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "provider-7/claude-sonnet-4-20250514",
    "max_tokens": 100,
    "stream": true,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

OpenAI Responses API (`/v1/responses`)

For GPT Codex models. Automatically adds provider-5/ prefix to model names.

Non-Streaming:

curl -X POST http://localhost:8787/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "gpt-5.1-codex",
    "input": "Say hello"
  }'

Streaming:

curl -X POST http://localhost:8787/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "gpt-5.1-codex",
    "input": "Say hello",
    "stream": true
  }'

Count Tokens

curl -X POST http://localhost:8787/v1/messages/count_tokens \
  -H "Content-Type: application/json" \
  -H "x-api-key: YOUR_API_KEY" \
  -d '{
    "model": "claude-sonnet-4-20250514",
    "messages": [{"role": "user", "content": "Hello, world!"}]
  }'

Response:

{"input_tokens": 4}

Configuration Constants

The following constants are defined in src/index.ts:

Constant	Value	Description
`A4F_BASE_URL`	`https://api.a4f.co/v1`	A4F API base URL
`A4F_PROVIDER_PREFIX`	`provider-7`	Prefix for Claude models
`A4F_RESPONSES_PROVIDER_PREFIX`	`provider-5`	Prefix for GPT-5.1 Codex models

Workarounds & Implementation Notes

Reasoning Summary Stripping

The /v1/responses endpoint strips the reasoning.summary field from requests before forwarding to A4F. This is because A4F's streaming implementation doesn't properly support this field. See handleResponses() for details.

Local Token Counting

Token counts are calculated locally using @lenml/tokenizer-claude rather than relying on A4F's response. This ensures consistent token counting across streaming and non-streaming requests. See countTokens().

Tool Choice Mapping

Anthropic's tool_choice.type: "any" maps to OpenAI's "required", not "any". See convertToolChoice().

Path Normalization

The proxy handles various path prefix issues that can occur with different clients:

Double /v1 prefix: /v1/v1/messages → /v1/messages
Double slash prefix: //v1/messages → /v1/messages
Flexible matching: Both /v1/messages and /messages work

See the main handler in src/index.ts for implementation details.

Local Development

Prerequisites

Bun runtime installed

Install Dependencies

bun install

Create Local Secrets File

Create a .dev.vars file (already in .gitignore):

# .dev.vars
A4F_API_KEY=your-real-a4f-api-key
VALID_API_KEYS=test-key-1,test-key-2

Run Locally

bun run dev
# Or: bunx wrangler dev

The worker will start on http://localhost:8787.

Type Checking

bun run typecheck

CloudFlare Worker

⚠️ Enabling/Disabling the Worker

IMPORTANT: For security and cost control, you should disable the worker when not in use.

Disable the Worker (Recommended when not in use)

Disabling stops all traffic to the worker but preserves your configuration and secrets.

Via CLI:

bunx wrangler disable

Via CloudFlare Dashboard:

Go to CloudFlare Dashboard
Navigate to Workers & Pages
Click on devkit-anthropic-proxy
Go to Settings
Click Disable

Enable the Worker

Via CLI:

bunx wrangler enable

Via CloudFlare Dashboard:

Go to CloudFlare Dashboard
Navigate to Workers & Pages
Click on devkit-anthropic-proxy
Go to Settings
Click Enable

Note: Disabling the worker is non-destructive. All secrets, configuration, and code are preserved. You can re-enable at any time.

Deployment

Prerequisites

CloudFlare account
Wrangler CLI (included via bunx)

Initial Setup

# Login to CloudFlare (first time only)
bunx wrangler login

# Set up required secrets
bunx wrangler secret put A4F_API_KEY
# When prompted, enter your A4F API key

bunx wrangler secret put VALID_API_KEYS
# When prompted, enter comma-separated keys: key1,key2,key3

# Deploy
bun run deploy

Deploy Updates

bun run deploy
# Or: bunx wrangler deploy

View Deployment Versions

Via CLI:

bunx wrangler deployments list

Via Dashboard:

Go to Workers & Pages → devkit-anthropic-proxy
Click on Deployments tab

Rollback to Previous Version

bunx wrangler rollback

Managing API Keys

This worker uses a dual-key authentication system:

A4F API Key (A4F_API_KEY): Your real A4F API key (never exposed to users)
User API Keys (VALID_API_KEYS): Keys you distribute to your users/clients

Update the Backend A4F Key

bunx wrangler secret put A4F_API_KEY
# When prompted, enter your A4F API key

Update User Keys

User keys are stored as a comma-separated list:

bunx wrangler secret put VALID_API_KEYS
# When prompted, enter: key1,key2,key3

Example input:

kL5vJ2fL3oO8cH3iO4lU5eT0uX9wP7rJ,aB2cD4eF6gH8iJ0kL2mN4oP6qR8sT0uV

List Current Secrets

bunx wrangler secret list

This shows secret names (not values) that are currently configured.

Delete a Secret

bunx wrangler secret delete SECRET_NAME

Generating Secure User Keys

Generate cryptographically secure random keys:

# Using openssl (32 hex characters)
openssl rand -hex 16

# Using openssl (longer key)
openssl rand -hex 32

# Using Node.js
node -e "console.log(require('crypto').randomBytes(16).toString('hex'))"

# Using Python
python3 -c "import secrets; print(secrets.token_hex(16))"

Monitoring & Logs

View Real-Time Logs

Stream live logs from the production worker:

bunx wrangler tail

With pretty formatting:

bunx wrangler tail --format=pretty

Filter by status:

# Only errors
bunx wrangler tail --status=error

# Only successful requests
bunx wrangler tail --status=ok

CloudFlare Dashboard Analytics

Go to CloudFlare Dashboard
Navigate to Workers & Pages → devkit-anthropic-proxy
View the Analytics tab for:
- Request counts
- Error rates
- CPU time usage
- Geographic distribution

Security Best Practices

Keep the worker disabled when not in use - This prevents unauthorized access and unexpected costs
Rotate user keys periodically - Generate new keys and update VALID_API_KEYS regularly
Use strong, random keys - Always use cryptographically secure random generators
Monitor usage via CloudFlare Dashboard - Check for unusual patterns or unauthorized access attempts
Never commit secrets - The .dev.vars file is gitignored; never put real keys in wrangler.toml
Limit key distribution - Only share user keys with trusted parties

Roo Code / Cline Configuration

To use this proxy with Roo Code v3.35.3+ or Cline, configure the Anthropic provider:

Step-by-Step Configuration

Open Settings → Go to Providers tab
Create a new Configuration Profile:
- Click the + button next to "Configuration Profile"
- Name it something like "A4F Claude Proxy"
Configure the settings:

Setting	Value
API Provider	`Anthropic`
Anthropic API Key	Your user API key (from `VALID_API_KEYS`)
Pass Anthropic API Key as Authorization header	✅ Checked
Use custom base URL	✅ Checked
Custom Base URL	`https://your-worker-name.your-subdomain.workers.dev` OR `http://localhost:8787`
Model	`claude-opus-4-5-20251101` (or any Claude model)
Tool Call Protocol	`XML` (recommended for A4F)

Important Notes

Model Name: Use the model name WITHOUT the provider prefix:
- ✅ Correct: claude-sonnet-4-20250514
- ❌ Wrong: provider-7/claude-sonnet-4-20250514
The proxy automatically adds the provider-7/ prefix when forwarding to A4F.
Tool Call Protocol: Set to XML because A4F doesn't support native function calling for Claude models.

Troubleshooting

"Missing API key" Error

Cause: No API key provided in the request.

Solution: Include either:

x-api-key: YOUR_KEY header (Anthropic style)
Authorization: Bearer YOUR_KEY header (OpenAI style)

"Invalid API key" Error

Cause: The provided key is not in the VALID_API_KEYS list.

Solution:

Verify the key is correct
Check that VALID_API_KEYS is set: bunx wrangler secret list
Update the keys if needed: bunx wrangler secret put VALID_API_KEYS

"Server configuration error: A4F API key not configured"

Cause: The A4F_API_KEY secret is not set.

Solution:

bunx wrangler secret put A4F_API_KEY

Worker Returns 404

Cause: The worker may be disabled or the endpoint doesn't exist.

Solution:

Enable the worker: bunx wrangler enable
Verify the endpoint path (e.g., /v1/messages not /messages)

Streaming Not Working

Cause: Client may not support SSE or connection is being buffered.

Solution:

Ensure stream: true is in the request body
Check that your client supports Server-Sent Events
Disable any response buffering in proxies/load balancers

Token Count Mismatch

Cause: Token counting uses @lenml/tokenizer-claude which may differ slightly from Anthropic's internal tokenizer.

Solution: This is expected behavior. The counts are accurate for billing estimation but may vary by a few tokens from Anthropic's official API.

Deployment Fails

Cause: Various issues with wrangler or CloudFlare account.

Solution:

Ensure you're logged in: bunx wrangler login
Check your CloudFlare account has Workers enabled
Verify wrangler.toml syntax is correct
Try: bunx wrangler deploy --dry-run to test without deploying

Architecture

User Request (Anthropic/OpenAI format)
        │
        ▼
┌─────────────────────────────┐
│   CloudFlare Worker         │
│   ┌─────────────────────┐   │
│   │ Validate User Key   │   │
│   └─────────────────────┘   │
│            │                │
│            ▼                │
│   ┌─────────────────────┐   │
│   │ Route by Endpoint   │   │
│   │ /v1/messages        │──►│ Convert Anthropic → OpenAI
│   │ /v1/chat/completions│──►│ Pass-through
│   │ /v1/responses       │──►│ Add provider-5 prefix
│   └─────────────────────┘   │
│            │                │
│            ▼                │
│   ┌─────────────────────┐   │
│   │ Forward to A4F      │   │
│   │ (with A4F_API_KEY)  │   │
│   └─────────────────────┘   │
│            │                │
│            ▼                │
│   ┌─────────────────────┐   │
│   │ Convert Response    │   │
│   │ (if needed)         │   │
│   └─────────────────────┘   │
└─────────────────────────────┘
        │
        ▼
User Response (matching request format)

Available Models

Models available via the /v1/models endpoint:

Claude Models (Anthropic)

claude-opus-4-5-20251101
claude-sonnet-4-5-20250929
claude-sonnet-4-20250514
claude-3-7-sonnet-20250219
claude-3-5-sonnet-20241022
claude-3-5-sonnet-20240620
claude-3-5-haiku-20241022
claude-haiku-4-5-20251001
claude-3-haiku-20240307

GPT Codex Models (OpenAI)

gpt-5-codex
gpt-5.1-codex
gpt-5.1-codex-mini
gpt-5.1-codex-max

Check A4F's documentation for the full list of available models.

License

ISC

Name		Name	Last commit message	Last commit date
Latest commit History 9 Commits
src		src
.gitignore		.gitignore
README.md		README.md
package.json		package.json
tsconfig.json		tsconfig.json
wrangler.toml		wrangler.toml

edxeth/a4f-proxy

Folders and files

Latest commit

History

Repository files navigation

Devkit Anthropic Proxy

Features

Anthropic API (Messages)

OpenAI API (Chat Completions)

OpenAI Responses API

Common Features

API Endpoints

Authentication

Model Name Handling

Health Check

List Models

Anthropic Messages API (/v1/messages)

OpenAI Chat Completions API (/v1/chat/completions)

OpenAI Responses API (/v1/responses)

Count Tokens

Configuration Constants

Workarounds & Implementation Notes

Reasoning Summary Stripping

Local Token Counting

Tool Choice Mapping

Path Normalization

Local Development

Prerequisites

Install Dependencies

Create Local Secrets File

Run Locally

Type Checking

CloudFlare Worker

⚠️ Enabling/Disabling the Worker

Disable the Worker (Recommended when not in use)

Enable the Worker

Deployment

Prerequisites

Initial Setup

Deploy Updates

View Deployment Versions

Rollback to Previous Version

Managing API Keys

Update the Backend A4F Key

Update User Keys

List Current Secrets

Delete a Secret

Generating Secure User Keys

Monitoring & Logs

View Real-Time Logs

CloudFlare Dashboard Analytics

Security Best Practices

Roo Code / Cline Configuration

Step-by-Step Configuration

Important Notes

Troubleshooting

"Missing API key" Error

"Invalid API key" Error

"Server configuration error: A4F API key not configured"

Worker Returns 404

Streaming Not Working

Token Count Mismatch

Deployment Fails

Architecture

Available Models

Claude Models (Anthropic)

GPT Codex Models (OpenAI)

License

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Anthropic Messages API (`/v1/messages`)

OpenAI Chat Completions API (`/v1/chat/completions`)

OpenAI Responses API (`/v1/responses`)

Packages