agentlens

One import. See where your LLM money goes. Get told exactly how to cut it.

AgentLens wraps the Anthropic and OpenAI Python SDKs to capture cost and performance telemetry — without changing how your code works or adding any latency.

Install

pip install useagentlens

Quickstart

Anthropic

# Before
from anthropic import Anthropic

# After — one line change
from agentlens import Anthropic

client = Anthropic(
    agentlens_key="al_live_...",   # from useagentlens.com/settings
)

# Everything else stays the same
msg = client.messages.create(
    model="claude-haiku-4-5-20251001",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)

OpenAI

# Before
from openai import OpenAI

# After — one line change
from agentlens import OpenAI

client = OpenAI(
    agentlens_key="al_live_...",   # from useagentlens.com/settings
)

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

Labelling requests

Tag your requests to group costs by feature, user session, or agent step:

client = Anthropic(
    agentlens_key="al_live_...",   # from useagentlens.com/settings
    endpoint="customer-support",      # groups costs by feature in the dashboard
    session_id=user_session_id,       # tracks cost across a full conversation
    step=step_number,                 # tracks cost growth across agent steps
)

# Or override per-call without creating a new client
step2_client = client.with_options(step=2)

How it works

• Your existing Anthropic/OpenAI API key is passed straight through — AgentLens never sees it
• After each API response, usage data is read and sent to the AgentLens ingest API on a background thread
• Zero latency added to your application
• If the AgentLens API is unreachable, events are silently dropped — your code always continues normally

Dashboard

Sign in at useagentlens.com to see:

• Total spend over time
• Cost breakdown by endpoint, model, and session
• P95 latency alongside average latency
• Actionable insights: system prompt bloat, model mismatch, agent loop explosion

Optional dependencies

Install only what you need:

pip install "useagentlens[anthropic]"   # Anthropic wrapper only
pip install "useagentlens[openai]"      # OpenAI wrapper only
pip install "useagentlens[all]"         # Both

Debug mode

No API key? Print events to stdout instead of sending them:

client = Anthropic(agentlens_debug=True)

License

MIT