SDK Diff: Anthropic vs OpenAI-compat (Ollama / OpenAI / Together / most open-source)

繁體中文 | 简体中文 | English

Same ReAct loop, 3 key differences between SDKs. Pairs with SKILL.md Step 3.

TL;DR comparison

Part	Anthropic SDK	OpenAI-compat SDK
Tool schema wrap	tools=[{name, description, input_schema}]	tools=[{"type": "function", "function": {name, description, parameters}}]
Schema field name	input_schema	parameters
Reading tool call	[b for b in resp.content if b.type == "tool_use"]	resp.choices[0].message.tool_calls
Args format	call.input is already a dict	call.function.arguments is a JSON string — needs json.loads(...)
Stop detection	resp.stop_reason == "end_turn"	resp.choices[0].finish_reason == "stop"
Assistant turn append	messages.append({"role": "assistant", "content": resp.content})	messages.append({"role": "assistant", "content": msg.content, "tool_calls": msg.tool_calls})
Tool result append	messages.append({"role": "user", "content": [{"type": "tool_result", "tool_use_id": call.id, "content": obs}]})	messages.append({"role": "tool", "tool_call_id": tc.id, "content": obs})
Exception class	anthropic.RateLimitError / anthropic.APIConnectionError	openai.RateLimitError / openai.APIConnectionError

Side-by-side single-turn tool call

Anthropic

import anthropic

client = anthropic.Anthropic()

TOOLS = [{
    "name": "get_weather",
    "description": "Use this when the user asks about current weather.",
    "input_schema": {
        "type": "object",
        "properties": {"city": {"type": "string"}},
        "required": ["city"]
    }
}]

resp = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=512,
    tools=TOOLS,
    messages=[{"role": "user", "content": "Is it raining in Taipei?"}]
)

# Read tool call
calls = [b for b in resp.content if b.type == "tool_use"]
city = calls[0].input["city"]   # already a dict

OpenAI-compat (Ollama)

from openai import OpenAI
import json

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")

TOOLS = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Use this when the user asks about current weather.",
        "parameters": {       # ← different name
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"]
        }
    }
}]

resp = client.chat.completions.create(
    model="qwen2.5:3b",
    tools=TOOLS,
    messages=[{"role": "user", "content": "Is it raining in Taipei?"}]
)

# Read tool call
tc = resp.choices[0].message.tool_calls[0]
args = json.loads(tc.function.arguments)   # ← needs json.loads
city = args["city"]

Side-by-side ReAct loop (multi-turn)

Anthropic full loop

messages = [{"role": "user", "content": "..."}]
for step in range(5):
    resp = client.messages.create(model=MODEL, max_tokens=1024, tools=TOOLS, messages=messages)
    messages.append({"role": "assistant", "content": resp.content})   # full content list appended
    if resp.stop_reason == "end_turn":
        break
    tool_results = []
    for call in [b for b in resp.content if b.type == "tool_use"]:
        obs = TOOL_IMPL[call.name](call.input)
        tool_results.append({"type": "tool_result", "tool_use_id": call.id, "content": obs})
    messages.append({"role": "user", "content": tool_results})   # tool results wrapped in user message

OpenAI-compat full loop

messages = [{"role": "user", "content": "..."}]
for step in range(5):
    resp = client.chat.completions.create(model=MODEL, tools=TOOLS, messages=messages)
    msg = resp.choices[0].message
    messages.append({"role": "assistant", "content": msg.content, "tool_calls": msg.tool_calls})
    if not msg.tool_calls:
        break
    for tc in msg.tool_calls:
        args = json.loads(tc.function.arguments)
        obs = TOOL_IMPL[tc.function.name](args)
        messages.append({"role": "tool", "tool_call_id": tc.id, "content": obs})   # dedicated role

4 easy-to-trip-on spots

1. parameters vs input_schema: most common copy-paste trap — pasting an Anthropic schema directly into OpenAI-compat silently fails (Ollama doesn't raise; it just doesn't call the tool).
2. call.input vs json.loads(arguments): forgetting json.loads on the OpenAI-compat side gives you a string instead of a dict — KeyError.
3. Different role for tool result: Anthropic uses role="user" + [{"type": "tool_result", ...}]; OpenAI-compat uses role="tool" + plain string content.
4. stop_reason vs finish_reason: both exist but with different field names and values. Anthropic "end_turn" / "tool_use"; OpenAI "stop" / "tool_calls".

Full comparison examples

Every Stage 3 exercise ships both starters:

• starter.py = OpenAI-compat / Ollama
• starter_anthropic.py = Anthropic

Compare any of ../../stage-3/02-multi-tool-selection/ ~ ../../stage-3/06-schema-design/.