How do you migrate from Anthropic to Plugsky?

Key facts

Anthropic endpoint	POST /v1/messages with an x-api-key header
Plugsky endpoint	POST https://api.plugsky.com/v1/chat/completions with Authorization: Bearer sk-live-…
System prompt	Anthropic's top-level system parameter becomes a system role message
Tool schema	input_schema becomes function.parameters in the OpenAI tools array
Streaming	content_block_delta events become choices[0].delta.content chunks
Model mapping	30+ models to map Claude tiers onto; verify plugsky-frontier, plugsky-pro or plugsky-lite on your evals
Feature gaps	Prompt-cache controls and extended-thinking blocks have no direct equivalent — strip them in the adapter
Product status	Live (chat, streaming, tools, JSON mode, embeddings); audio, images, batch and fine-tuning are coming soon

TL;DR

• The messages array is similar; the system prompt moves inside it.
• Tools need a wrapper: input_schema becomes function.parameters.
• Streaming event names differ but the token stream is equivalent.
• max_tokens is required on Anthropic and optional on Plugsky.
• Prompt caching and extended thinking have no direct equivalent — plan around them.

How it works, step by step

1. Inventory every Anthropic call site and note which features it uses (tools, vision, caching, thinking).
2. Create a Plugsky API key and confirm access with a single chat completion.
3. Map Claude models to Plugsky models and write an adapter that converts messages and tools.
4. Replace the Anthropic SDK with the OpenAI SDK pointed at https://api.plugsky.com/v1.
5. Convert streaming consumers from content_block_delta to delta.content chunks.
6. Run your eval suite against both providers and diff the outputs.
7. Cut over traffic gradually and keep the adapter reversible for one release cycle.

Original data

Try it yourself

Open the OpenAI migration checker →

How the Anthropic API differs

Four differences cause most migration bugs:

• Auth: Anthropic sends x-api-key: sk-ant-…; Plugsky sends Authorization: Bearer sk-live-….
• System prompt: Anthropic takes a top-level system string; OpenAI-compatible APIs use a message with role: "system".
• Tools: Anthropic tools use input_schema; Plugsky uses function.parameters inside a {"type": "function", "function": {…}} wrapper.
• Streaming: Anthropic emits named events such as content_block_delta with text_delta; Plugsky emits OpenAI chunks with choices[0].delta.content.

Everything else — multi-turn history, temperature, stop sequences, tool result turns — maps directly.

Step-by-step conversion

An Anthropic call:

import anthropic

client = anthropic.Anthropic(api_key="sk-ant-…")
msg = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    system="You are a terse support agent.",
    messages=[{"role": "user", "content": "How do refunds work?"}],
)

The Plugsky equivalent:

from openai import OpenAI

client = OpenAI(api_key="sk-live-…", base_url="https://api.plugsky.com/v1")
resp = client.chat.completions.create(
    model="plugsky-pro",
    max_tokens=1024,
    messages=[
        {"role": "system", "content": "You are a terse support agent."},
        {"role": "user", "content": "How do refunds work?"},
    ],
)
print(resp.choices[0].message.content)

Keep an adapter function so the mapping lives in one place. It should accept Anthropic-shaped input and return an OpenAI-shaped request; that lets you A/B both providers during migration.

Converting tools and streaming

Tool conversion is mechanical. Anthropic:

{"name": "get_weather", "description": "Get weather",
 "input_schema": {"type": "object", "properties": {"city": {"type": "string"}}}}

Plugsky and other OpenAI-compatible endpoints:

{"type": "function",
 "function": {"name": "get_weather", "description": "Get weather",
              "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}}

Tool choice moves from {"type": "tool", "name": "get_weather"} to {"type": "function", "function": {"name": "get_weather"}}. For streaming, replace handlers that listen for content_block_delta with a loop over chunks and append chunk.choices[0].delta.content when it is present. Anthropic's content_block_start and message_delta events have no direct equivalents, so usage accounting moves to the final chunk or a non-streaming call.

What to validate before cutover

Run your eval suite on both providers with identical prompts. Score three things: task success, tool-call correctness and output format. Watch for workloads that depend on features Plugsky does not replace today — prompt caching and extended thinking are the two that force architectural changes, because the cache-control blocks must be stripped before the request is sent.

For high-volume workloads, route cheap classification steps and hard reasoning steps to different Plugsky models instead of one Claude tier. plugsky-lite handles extraction and routing; plugsky-pro or plugsky-frontier handle synthesis. Because Plugsky is OpenAI-compatible, you can keep the adapter in place and switch providers by changing one base URL if an eval regresses.

Honest comparison

Capability	Plugsky (OpenAI-compatible)	Anthropic Messages API	Dual-provider router
Auth	Bearer sk-live-…	x-api-key header	Two credential types
System prompt	system role message	Top-level system field	Adapter maps both
Tools	functions in the tools array	input_schema in the tools array	Translate schemas at the boundary
Streaming	choices[].delta.content	content_block_delta events	Normalize both formats
Unique features	Flat plans, fusion model, residency options	Prompt caching, extended thinking	Keep both providers for now
Migration effort	Rewrite calls to the OpenAI SDK plus an adapter	Incumbent	Ongoing maintenance burden

Frequently asked questions

Can I keep using the Anthropic SDK?

No. The Messages API is not OpenAI-compatible, so you switch to an OpenAI-compatible client and convert request shapes in an adapter.

Where does the system prompt go?

Move the top-level system string into the first message with role: "system" in the messages array.

How do I convert tools?

Wrap each tool as {"type": "function", "function": {...}} and rename input_schema to parameters. Tool choice uses the same wrapper.

How does streaming change?

Read chunk.choices[0].delta.content instead of listening for content_block_delta events. Usage moves to the final chunk.

Is max_tokens required?

No. Anthropic requires max_tokens; on Plugsky it is optional, though setting it explicitly helps control cost and latency.

What Claude features do not carry over?

Prompt-cache controls and extended-thinking blocks have no direct equivalent today. Strip them in the adapter and re-evaluate those workloads.

How should I pick Plugsky models?

Map Claude tiers to plugsky-pro or plugsky-frontier for complex work and plugsky-lite for high-volume steps, then validate on your evals.

Can I run both providers side by side?

Yes. Keep a provider switch behind one adapter and route a small share of traffic to Plugsky while you compare quality and cost.