tracellm

Lightweight tracing for LLM applications. One decorator — every API interaction logged locally, queryable from your terminal.

No backend. No signup. Nothing leaves your machine.

Install

pip install tracellm

Usage

from tracellm import trace
import groq

client = groq.Groq(api_key="your-key")

@trace
def ask(model, messages):
    return client.chat.completions.create(model=model, messages=messages)

ask(
    model="llama-3.1-8b-instant",
    messages=[{"role": "user", "content": "Explain black holes in one line"}]
)

Every call is traced automatically. No try/except. No setup.

Query traces

python -m tracellm.cli --Status failed
python -m tracellm.cli --Latency 2.0
python -m tracellm.cli --Model llama-3.1-8b-instant
python -m tracellm.cli --Status failed --Latency 1.5
python -m tracellm.cli --Time "2026-04-03"

Cost tracking

# cost per trace
python -m tracellm.cli --Cost

# full summary by model
python -m tracellm.cli --Cost Summary

Output: === Cost Summary === llama-3.1-8b-instant Calls : 8 Tokens : 405 Cost : $0.000020

Total calls made : 8 Total tokens used: 405 Total cost : $0.000020

What gets captured

• Model, prompt, response
• Tokens used, latency, finish reason
• Error type and message on failures
• Timestamp for every call

Pricing

Default pricing is bundled. To override, create ~/.tracellm/pricing.json:

{
  "my-custom-model": 0.05
}

Values are per million tokens.

Limitations

Storage is append-only JSON lines. Latency filter supports >=, exact match for everything else. Early days.

Roadmap

• Binary storage for faster querying at scale
• Async tracing support
• Terminal dashboard