tokenguard (Python SDK)

Monitor AI agents, track LLM costs, debug failures — in 2 lines of code.

Install

pip install tokenguard

Quick Start — Auto-wrap Groq (zero code change)

from groq import Groq
from tokenguard import TokenGuard

aw = TokenGuard(
    api_key="tg_live_...",           # from your dashboard Settings
    base_url="http://localhost:3000", # your TokenGuard URL
)

# Wrap your existing Groq client — ONE LINE
groq = aw.wrap_groq(Groq(api_key="..."), agent_name="DietAgent")

# Use exactly as before — tracking happens automatically
response = groq.chat.completions.create(
    model="llama3-8b-8192",
    messages=[{"role": "user", "content": "Give me a low-carb meal plan"}],
)

print(response.choices[0].message.content)
# → Your dashboard now shows cost, tokens, latency for this call

Quick Start — Auto-wrap OpenAI

from openai import OpenAI
from tokenguard import TokenGuard

aw = TokenGuard(api_key="tg_live_...", base_url="http://localhost:3000")

# Wrap your OpenAI client
openai = aw.wrap_openai(OpenAI(api_key="..."), agent_name="SupportBot")

response = openai.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello"}],
)

Manual Tracing (full control)

from tokenguard import TokenGuard

aw = TokenGuard(api_key="tg_live_...", base_url="http://localhost:3000")

def get_diet_plan(user_message):
    with aw.trace("DietSuggestionAgent") as trace:

        # Track the Groq call
        llm = trace.llm(model="llama3-8b-8192", provider="groq")
        response = groq_client.chat.completions.create(
            model="llama3-8b-8192",
            messages=[{"role": "user", "content": user_message}],
        )
        llm.end(
            input_tokens=response.usage.prompt_tokens,
            output_tokens=response.usage.completion_tokens,
        )

        # Track a tool call (optional)
        tool = trace.tool("nutrition_database_lookup")
        foods = lookup_foods(user_message)
        tool.end(result=foods)

        return response.choices[0].message.content

What appears in your dashboard

Every call shows:

• 💰 Cost — exact cost per request
• ⏱️ Latency — how long it took
• 🔢 Tokens — input and output token counts
• 🐛 Errors — full error details with stack traces
• 📊 Agent breakdown — which agents cost the most

Supported providers

Before you exit your script

aw.flush()  # makes sure all traces are sent before process ends