Using prxy.monster with the Anthropic SDK

prxy.monster speaks the native Anthropic API at https://api.prxy.monster — same /v1/messages shape, same headers, same streaming format. Set one env var, your existing @anthropic-ai/sdk (or anthropic Python) code routes through prxy.monster unchanged.

Install

npm install @anthropic-ai/sdk
# or
pip install anthropic

Configure

export ANTHROPIC_BASE_URL=https://api.prxy.monster
export ANTHROPIC_API_KEY=prxy_live_xxxxxxxxxxxxxxxxxxxxxxxx

Code change

None.

// Before AND after — no diff
import Anthropic from '@anthropic-ai/sdk';
 
const client = new Anthropic(); // reads env
 
const msg = await client.messages.create({
  model: 'claude-sonnet-4-6',
  max_tokens: 256,
  messages: [{ role: 'user', content: 'hi' }],
});

const client = new Anthropic({
  baseURL: 'https://api.prxy.monster',
  apiKey: process.env.ANTHROPIC_API_KEY,
});

# Before AND after — no diff
from anthropic import Anthropic
 
client = Anthropic()  # reads env
 
msg = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=256,
    messages=[{"role": "user", "content": "hi"}],
)

from anthropic import Anthropic
 
client = Anthropic(
    base_url="https://api.prxy.monster",
    api_key="prxy_live_xxxxxxxxxxxxxxxxxxxxxxxx",
)

Verify

curl https://api.prxy.monster/health

Or:

prxy doctor

What you get

• Infinite context — long Claude conversations stop hitting the context wall.
• MCP optimization — if you wire MCP servers through Claude tool-use, irrelevant tool defs get pruned per request (~67k → ~8k tokens on a real Claude Code session).
• Semantic cache — similar prompts return cached responses.
• Pattern memory — the issue was X, the fix is Y style insights get logged and re-injected.

Recommended pipeline

For long-running Claude apps or coding agents:

PRXY_PIPE=mcp-optimizer,semantic-cache,patterns,ipc

For Claude Code specifically, add the compaction bridge:

PRXY_PIPE=mcp-optimizer,compaction-bridge,rehydrator,semantic-cache,patterns,ipc

Streaming

Anthropic’s per-block SSE format passes through unchanged — message_start, content_block_start, content_block_delta, message_stop all flow exactly as Anthropic emits them.

const stream = client.messages.stream({
  model: 'claude-sonnet-4-6',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'write a haiku' }],
});
 
for await (const event of stream) {
  if (event.type === 'content_block_delta' && event.delta.type === 'text_delta') {
    process.stdout.write(event.delta.text);
  }
}

Cache hits are replayed as synthetic Anthropic SSE so your stream parser doesn’t notice the difference.

Common issues

• Vertex / Bedrock SDK variants — @anthropic-ai/vertex-sdk and @anthropic-ai/bedrock-sdk use cloud-provider auth, so they don’t talk to prxy.monster directly. Use the standard @anthropic-ai/sdk and route Bedrock models via model: "bedrock/<model-id>" (see Bedrock provider).
• Beta headers (anthropic-beta: prompt-caching-2024-07-31 etc.) — pass-through. We don’t strip them.
• Tool use / input_schema — pass-through. mcp-optimizer reads tool defs to score relevance; tool execution still happens client-side as usual.

Full example

Plain Node script: github.com/Ekkos-Technologies-Inc/prxy-examples/tree/main/examples/anthropic-quickstart