Anthropic

One patch call. Every messages.create becomes an LLM span — sync, async, streaming.

1. Get an API key

export STASO_API_KEY=ak_...

2. Install

pip install "staso[anthropic]"

Python 3.11+.

3. Patch and go

import staso as st
from staso.integrations import patch_anthropic
import anthropic

st.init(agent_name="my-agent")
patch_anthropic()

client = anthropic.Anthropic()
client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=256,
    messages=[{"role": "user", "content": "hi"}],
)

Open the Staso dashboard — the call shows up as an LLM span with model, tokens, and latency.

What it captures

Tokens (input, output, cache read, cache creation).
Request params (model, temperature, max_tokens, top_p, stop_sequences, tool_choice, thinking).
Response (content blocks, tool use, thinking, stop_reason, response_id).
Streaming (stream=True and messages.stream(...) both wrapped).

Guard

Guard evaluates every tool_use block in the response. Blocked tools raise staso.GuardBlocked; modified tools have their input rewritten in place. See error handling.

Tool spans

Wrap your tools with @st.tool to capture their inputs and outputs as child spans:

@st.tool
def search_faq(query: str) -> str:
    return db.search(query)

See Python SDK quickstart for the full set of decorators.