Trace LiteLLM calls with Future AGI

Trace LiteLLM Python SDK calls in Observe. For a Proxy server, see LiteLLM Proxy tracing.

Install

Use Python 3.10–3.13. The traceAI-litellm package declares LiteLLM 1.43.0 or later as a dependency.

pip install 'traceAI-litellm==0.1.10' 'litellm>=1.43.0'

Set credentials and capture preferences

Set these in your shell, then run the Python blocks below in one script. Choose a model available to your OpenAI account; LiteLLM’s openai/ prefix selects that provider.

export FI_API_KEY="your-futureagi-api-key"
export FI_SECRET_KEY="your-futureagi-secret-key"
export OPENAI_API_KEY="your-openai-api-key"
export LITELLM_MODEL="openai/your-available-model"

Instrumentation can send messages, model responses and error details to Future AGI. To hide input and output values and messages, set FI_HIDE_INPUTS=true and FI_HIDE_OUTPUTS=true before initialization. See mask span attributes. These switches are not a general scrubber for custom attributes or exception text; avoid secrets in request data and error messages.

Register the trace provider

Name the Observe project that will receive spans.

from fi_instrumentation import register
from fi_instrumentation.fi_types import ProjectType

trace_provider = register(
    project_type=ProjectType.OBSERVE,
    project_name="litellm_project",
)

Instrument LiteLLM

Initialize instrumentation before making requests.

from traceai_litellm import LiteLLMInstrumentor

LiteLLMInstrumentor().instrument(tracer_provider=trace_provider)

Run a request

import os
import litellm

try:
    response = litellm.completion(
        model=os.environ["LITELLM_MODEL"],
        messages=[{"role": "user", "content": "What's the capital of India?"}],
    )
    print(response.choices[0].message.content)
finally:
    trace_provider.shutdown()

shutdown() sends pending spans before this short-lived process exits.

View the trace

Open Observe in Future AGI and select litellm_project to inspect the request’s input, output, timing and available model usage. Usage fields depend on the response returned by the model provider.