nolma

AI gateway SDK — cost control, session tracing, and security for LLM agents.

One URL change. See every dollar your AI agents spend.

Installation

pip install nolma

Quick start

from nolma import Nolma

nolma = Nolma(api_key="nm_live_abc123")

@nolma.session(name="support-bot", user_id="user_42")
async def my_agent(query: str):
    # Your LLM calls go here — Nolma tracks everything
    ...

Session decorator

Wraps your agent function with automatic session tracking via contextvars:

@nolma.session(name="research-agent", user_id="user_123")
async def research(topic: str):
    result = await llm_call(topic)
    return result

Context manager

async with nolma.start_session(name="email-drafter") as ctx:
    print(ctx.session_id)  # sess_abc123...
    result = await draft_email(input)

Header injection

Pass Nolma headers to any LLM SDK:

import anthropic

client = anthropic.Anthropic(
    base_url=nolma.gateway_url + "/anthropic",
    default_headers=nolma.get_headers(),
)

Signals (Nolma Lens)

Track what users do with LLM outputs:

nolma.signal(session_id, action="accepted")
nolma.signal(session_id, action="edited", edit_distance=42)
nolma.signal(session_id, action="regenerated")

Context scanning (Nolma Shield)

Detect injection attacks in RAG content:

result = await nolma.scan_context(
    content=retrieved_doc,
    source="rag",
    agent="support-bot",
)
if result["decision"] == "BLOCK":
    raise SecurityError("Injection detected")

Documentation

Full docs at docs.nolma.ai

License

MIT