LLM Providers

Every provider is auto-instrumented by init() — construct and use the client as normal.

All providers below are auto-instrumented by `init()` — just construct and use the client as normal. The manual patch_* / patch* call is shown for completeness (use it for a client created before init(), or in async/edge setups).

Anthropic (Claude)

import evalkit, anthropic
evalkit.init(subscription_key="tk_live_...", service_name="claude-app")

client = anthropic.Anthropic()                # auto-traced
resp = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Explain tracing in one line."}],
)

# Async
from anthropic import AsyncAnthropic
aclient = AsyncAnthropic()                     # auto-traced
# Manual (optional): client created before init / custom wrapper
evalkit.patch_anthropic_client(client)
evalkit.patch_async_anthropic_client(aclient)

Captures: model, prompt, completion, input/output tokens, stop reason, streaming deltas, and tool use blocks — each tool the model requests is attached to the call, no manual spans needed.

Claude via Amazon Bedrock

import boto3, evalkit
bedrock = boto3.client("bedrock-runtime")
evalkit.patch_bedrock_client(bedrock)         # then invoke_model / converse as usual

Claude via Google Vertex

from anthropic import AnthropicVertex
v = AnthropicVertex(region="us-east5", project_id="my-proj")
evalkit.patch_anthropic_vertex_client(v)

OpenAI

import evalkit, openai
evalkit.init(subscription_key="tk_live_...")
client = openai.OpenAI()                       # auto-traced (sync)
# async: openai.AsyncOpenAI() auto-traced
# manual: evalkit.patch_openai_client(client) / evalkit.patch_async_openai_client(aclient)

OpenAI-compatible providers (Groq, Together, Fireworks, xAI, …): point the OpenAI client at their base_url — they trace through the OpenAI instrumentation automatically. In Python you can also route them through LiteLLM (below).

Google Gemini

# New unified SDK (google-genai) — also used by Google ADK
import evalkit
evalkit.init(subscription_key="tk_live_...")
evalkit.patch_google_genai()                   # patches google.genai Models
# Legacy google-generativeai:
#   evalkit.patch_google_ai_model(model)

Cohere

evalkit.patch_cohere_client(cohere_client)

Mistral (Python)

from mistralai import Mistral
client = Mistral(api_key=...)
evalkit.patch_mistral_client(client)           # async: patch_async_mistral_client(aclient)

LiteLLM (Python)

import evalkit, litellm
evalkit.init(subscription_key="tk_live_...")   # litellm.completion/acompletion auto-traced
# manual: evalkit.patch_litellm()
resp = litellm.completion(model="claude-opus-4-8", messages=[{"role":"user","content":"hi"}])

Support matrix

Provider	Python	TypeScript
OpenAI (sync/async)	✅	✅
Anthropic / Claude (sync/async)	✅	✅
Claude via Bedrock	✅	✅
Claude via Vertex	✅	✅
Google Gemini (genai / legacy / Vertex)	✅	✅
Cohere	✅	✅
Mistral	✅	—
LiteLLM	✅	—
Groq / OpenAI-compatible	✅ (OpenAI client / LiteLLM)	✅ (OpenAI client)