Harness SDK (Python)

Instrument a Python application that calls a model SDK directly (LiteLLM, OpenAI, Anthropic, or Google GenAI) to emit GenAI traces to Cloud & AI Cost Management.

Use the Harness SDK when your Python application calls a model provider SDK directly (LiteLLM, OpenAI, Anthropic, or Google GenAI), with no orchestration framework, and does not yet emit GenAI OpenTelemetry traces. You add two lines of instrumentation and the SDK emits spans with GenAI semantic conventions.

If you came from the quickstart's Instrument your application path, the snippet there is the short version. This page is the full setup: install, configure the endpoint, add instrumentation, and the per-client examples for LiteLLM, OpenAI, Anthropic, and Google GenAI.

Before you start, generate an ingestion token and connect a provider. Go to the AI Cost Management Quickstart for those steps.

GENAI SEMANTIC CONVENTIONS REQUIRED

Cost is calculated from OpenTelemetry traces with GenAI semantic conventions. Standard HTTP, database, or function spans do not carry the model name or token counts CACM needs. Go to how GenAI traces differ from standard OpenTelemetry to understand why, or to the GenAI Span Attribute Reference to review the attributes CACM reads.

---

How Instrumentation Works

GenAI instrumentation has three parts: the OpenTelemetry SDK, the GenAI instrumentation libraries, and the OTLP exporter. The Harness SDK bundles all three, so the two lines below turn all of them on at once. Go to the three parts of instrumentation to understand what each part does.

By default, the SDK traces none of your AI providers. Enable each provider your application uses (OpenAI, Anthropic, Google GenAI, or LiteLLM) with a single environment variable, so Harness captures cost data only for the calls you want to track. Go to Configure the SDK to set the correct environment variable for each provider.

If your calls run through an orchestration framework such as LangChain or Google ADK, use the matching open-source SDK integration instead, so the full workflow is captured.

---

Install the SDK

Install the SDK with the extra for your LLM client.

# For LiteLLM <a href="#for-litellm" id="for-litellm"></a>
pip install "harness-sdk[litellm]"

# For Anthropic Python client <a href="#for-anthropic-python-client" id="for-anthropic-python-client"></a>
pip install "harness-sdk[anthropic]"

# For OpenAI SDK <a href="#for-openai-sdk" id="for-openai-sdk"></a>
pip install "harness-sdk[openai]"

# For Google GenAI / Gemini / Vertex AI <a href="#for-google-genai-gemini-vertex-ai" id="for-google-genai-gemini-vertex-ai"></a>
pip install "harness-sdk[google-genai]"

The Anthropic extra requires additional packages. Install them separately if they are not already present:

pip install opentelemetry-instrumentation-anthropic opentelemetry-util-genai

If your application uses more than one provider, combine the extras in a single install with a comma-separated list:

pip install "harness-sdk[anthropic,google-genai,openai]"

---

Configure the SDK

Set these environment variables to point the SDK at the Harness endpoint. Replace <ACCOUNT_ID> with your account identifier (find it in the Harness URL when logged in) and <YOUR_TOKEN> with the ingestion token.

export HARNESS_SERVICE_NAME="my-ai-service"
export HARNESS_REPORTING_ENDPOINT="https://app.harness.io/udp-ingest/otel/v1/traces?accountIdentifier=<ACCOUNT_ID>&routingId=<ACCOUNT_ID>"
export HARNESS_REPORTING_TRACE_REPORTER_TYPE=OTLP_HTTP
export HARNESS_REPORTING_TOKEN="<YOUR_TOKEN>"

# Enable only the providers your application uses. <a href="#enable-only-the-providers-your-application-uses" id="enable-only-the-providers-your-application-uses"></a>
export HARNESS_ENABLE_AI_OPENAI=true
export HARNESS_ENABLE_AI_ANTHROPIC=true
export HARNESS_ENABLE_AI_GOOGLE_GENAI=true
export HARNESS_ENABLE_AI_LITELLM=true

Variable	Purpose
HARNESS_SERVICE_NAME	Service name that appears in Cost Explorer (example: customer-support-bot).
HARNESS_REPORTING_ENDPOINT	Harness UDP ingest endpoint with account ID.
HARNESS_REPORTING_TRACE_REPORTER_TYPE	Transport type. Use OTLP_HTTP.
HARNESS_REPORTING_TOKEN	The ingestion token generated in the Quickstart.

Enable per-provider instrumentation

Nothing is instrumented by default. Set the flag for each provider your application uses. Each flag must be exactly true.

Flag	What it traces
HARNESS_ENABLE_AI_OPENAI=true	OpenAI SDK
HARNESS_ENABLE_AI_ANTHROPIC=true	Anthropic SDK (including Vertex)
HARNESS_ENABLE_AI_GOOGLE_GENAI=true	Google GenAI / Gemini / Vertex
HARNESS_ENABLE_AI_LITELLM=true	LiteLLM

---

Add Instrumentation

Add these two lines at the start of the application, before importing any AI library:

from harness_sdk.agent import Agent
Agent().instrument()

CALL INSTRUMENT() BEFORE IMPORTING AI LIBRARIES

The SDK patches LLM libraries (LiteLLM, OpenAI, Anthropic, Google GenAI) and web frameworks (FastAPI, Flask) at import time via monkey-patching. Python executes module code once, so if import openai runs before Agent().instrument(), the client is never patched and no spans are emitted.

Select the client your application uses:

LiteLLM

Use case: Application uses LiteLLM to call multiple LLM providers (OpenAI, Anthropic, Bedrock, etc.).

Before you run the application, enable LiteLLM tracing in your environment. Without this, the SDK emits no spans.

export HARNESS_ENABLE_AI_LITELLM=true

With the provider enabled, instrument the application before it makes any model call. Every call is then traced automatically:

from harness_sdk.agent import Agent
Agent().instrument()  # Must be called before importing AI libraries

import litellm

# Every litellm.completion() call now emits a span with gen_ai.* attributes <a href="#every-litellmcompletion-call-now-emits-a-span-with-genai-attributes" id="every-litellmcompletion-call-now-emits-a-span-with-genai-attributes"></a>
resp = litellm.completion(
    model="anthropic/claude-3-5-sonnet-20241022",
    messages=[{"role": "user", "content": "Reply with one short sentence."}],
    max_tokens=64,
)
print(resp.choices[0].message.content)

What this produces:

• One span per litellm.completion() call.

• Span attributes: gen_ai.system=anthropic, gen_ai.request.model=claude-3-5-sonnet-20241022, gen_ai.usage.input_tokens, gen_ai.usage.output_tokens.

• Cost calculated from token counts and Anthropic pricing.

Anthropic Python client

Use case: Application uses the Anthropic Python client directly (not LiteLLM).

Before you run the application, enable Anthropic tracing in your environment. Without this, the SDK emits no spans.

export HARNESS_ENABLE_AI_ANTHROPIC=true

With the provider enabled, instrument the application before it makes any model call. Every call is then traced automatically:

from harness_sdk.agent import Agent
Agent().instrument()  # Must be called before importing AI libraries

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from env
msg = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=64,
    messages=[{"role": "user", "content": "Reply with one short sentence."}],
)
print(msg.content[0].text)

What this produces:

• One span per client.messages.create() call.

• Span attributes: gen_ai.system=anthropic, gen_ai.request.model=claude-3-5-sonnet-20241022, token usage.

• Cost calculated from token counts and Anthropic pricing.

OpenAI SDK

Use case: Application uses the OpenAI SDK directly.

Before you run the application, enable OpenAI tracing in your environment. Without this, the SDK emits no spans.

export HARNESS_ENABLE_AI_OPENAI=true

With the provider enabled, instrument the application before it makes any model call. Every call is then traced automatically:

from harness_sdk.agent import Agent
Agent().instrument()  # Must be called before importing AI libraries

import openai

client = openai.OpenAI()  # reads OPENAI_API_KEY from env
resp = client.chat.completions.create(
    model="gpt-4-turbo",
    messages=[{"role": "user", "content": "Reply with one short sentence."}],
    max_tokens=64,
)
print(resp.choices[0].message.content)

What this produces:

• One span per client.chat.completions.create() call.

• Span attributes: gen_ai.system=openai, gen_ai.request.model=gpt-4-turbo, token usage.

• Cost calculated from token counts and OpenAI pricing.

Google GenAI / Gemini

Use case: Application uses the Google GenAI SDK to call Gemini, either through the Gemini API or Vertex AI.

Before you run the application, enable Google GenAI tracing in your environment. Without this, the SDK emits no spans.

export HARNESS_ENABLE_AI_GOOGLE_GENAI=true

With the provider enabled, instrument the application before it makes any model call. Every call is then traced automatically:

from harness_sdk.agent import Agent
Agent().instrument()  # Must be called before importing AI libraries

from google import genai

client = genai.Client(
    vertexai=True,
    project="my-gcp-project",
    location="us-central1",
)
resp = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Reply with one short sentence.",
)
print(resp.text)

What this produces:

• One span per client.models.generate_content() call.

• Span attributes: gen_ai.system=gcp.vertex_ai, gen_ai.request.model=gemini-2.5-flash, token usage.

• Cost calculated from token counts and Gemini pricing.

---

Enrich Trace Data with Custom Attributes

By default, each span records the model, the token counts, and the cost of a call. This adds up to one combined cost figure, for example your total Gemini spend for the month, but with no breakdown of which customer or feature drove it. So when the bill jumps, you cannot tell what caused it.

Custom attributes solve this. You tag each call with the context you want to track, such as the customer or the feature, so you can answer questions such as "how much is each customer costing me?" or "which feature is the most expensive?"

For example, if one application serves many customers from the same code (a multi-tenant application), you can tag every call with the customer it belongs to. The example below adds two attributes:

• tenant.id: which customer made the call, so you can see cost per customer.

• agent.action.type: what the call was doing, such as generating an answer, so you can see cost per feature.

An attribute attaches to whichever span is active at the time, so you wrap the model call in a span and set the attributes inside it:

from google import genai
from opentelemetry import trace
from harness_sdk import set_span_attributes

client = genai.Client(
    vertexai=True,
    project="my-gcp-project",
    location="us-central1",
)
tracer = trace.get_tracer(__name__)

def generate_answer(tenant_id: str, prompt: str) -> str:
    with tracer.start_as_current_span("tenant.vertex.request"):
        set_span_attributes({
            "tenant.id": tenant_id,
            "agent.action.type": "generate-answer",
        })
        response = client.models.generate_content(
            model="gemini-2.5-flash",
            contents=prompt,
        )
        return response.text

Every span this function produces now carries the customer ID and the action type. In Cost Explorer, you can then group or filter your AI cost by customer or by feature, instead of seeing only a single combined total.

---

Reduce Trace Data Volume

Large prompt and response payloads inflate span volume and storage cost. To keep trace data manageable in high-traffic production:

• Disable payload capture (Harness SDK only): Set HARNESS_GEN_AI_PAYLOAD_CAPTURE_ENABLED=false so raw prompt and response text is not stored on spans.

• Scope instrumentation to LLM calls: Instrument the model calls that carry cost, not every function.

• Sample a percentage of traces: Export a representative sample rather than every trace.

---

Verify Traces in Cost Explorer

1. Run the application and trigger an LLM call.

2. Wait for traces to appear. They usually land within a few minutes; allow up to about 20 minutes.

3. Go to Cloud & AI Cost Management > Cost Explorer.

4. Select the AI Traces view or group by Service Name, and find your service (from HARNESS_SERVICE_NAME).

5. Select a service row to open the Service Traces drawer, and drill from a session down to the exact LLM call that drove the cost.

---

Debug Instrumentation Locally

When traces do not appear in Cost Explorer, it is hard to tell where the problem is, because the SDK sends spans straight to Harness where you cannot see them. These two environment variables make the SDK show its work on your own machine, so you can see a span created before it leaves your application:

Variable	What it does
HARNESS_ENABLE_CONSOLE_SPAN_EXPORTER=true	Prints a copy of every span to your terminal the moment it is created, so you can confirm the SDK is capturing your model calls.
HARNESS_LOG_LEVEL=DEBUG	Prints detailed SDK activity, including which libraries it instruments, which providers you enabled, and each attempt to send spans to Harness along with the response.

Add both to your existing setup, then run the application and trigger one model call:

# Add these to your existing configuration and provider flags <a href="#add-these-to-your-existing-configuration-and-provider-flags" id="add-these-to-your-existing-configuration-and-provider-flags"></a>
export HARNESS_ENABLE_CONSOLE_SPAN_EXPORTER=true
export HARNESS_LOG_LEVEL=DEBUG

Then match the terminal output to find the cause:

What you see	What it means	What to do
A span prints with gen_ai.* attributes	Instrumentation works. Your calls are captured, so the problem is getting them to Harness.	Go to Troubleshooting to fix the connection to Harness.
No span prints	The SDK is not capturing your calls. Usually the instrumentation call Agent().instrument() runs after the AI library is imported, or the provider flag for your client is not set to true.	Go to Add Instrumentation to fix the order, and Enable per-provider instrumentation to set the flag.

Turn both variables off in production. They add noise and overhead, and printing spans to the console is only useful while you are debugging.

---

Troubleshooting

The Harness SDK logs a 403 Forbidden response when exporting spans, and no traces reach Harness

A 403 means the ingestion token is invalid, expired, or lacks permission. Regenerate the token, set it as HARNESS_REPORTING_TOKEN, and confirm HARNESS_REPORTING_ENDPOINT includes the correct account identifier. If the error persists, contact Harness Support with your account ID and service account details.

The application calls Claude through the Anthropic SDK but no span appears in Cost Explorer

The SDK instruments no providers by default. Set HARNESS_ENABLE_AI_ANTHROPIC=true before running the application, and confirm Agent().instrument() runs before the Anthropic client is imported. To check locally, set HARNESS_ENABLE_CONSOLE_SPAN_EXPORTER=true and verify a span with gen_ai.* attributes prints to the terminal.

The application calls Gemini through the Google GenAI SDK but no span appears in Cost Explorer

The SDK instruments no providers by default. Set HARNESS_ENABLE_AI_GOOGLE_GENAI=true before running the application, and confirm Agent().instrument() runs before the Google GenAI client is imported. To check locally, set HARNESS_ENABLE_CONSOLE_SPAN_EXPORTER=true and verify a span with gen_ai.* attributes prints to the terminal.

---

Next Steps

• Go to the GenAI Span Attribute Reference to review the exact attributes CACM reads.

• Go to How AI traces work to understand trace attribution in depth.

• Go to AI Cost Troubleshooting if traces do not appear or show no cost.