Open-source Python SDK

Know what your agents actually did.

A flight recorder for AI agents. Every tool call, every LLM call, every error — captured as replayable evidence.

Caja negra para tus agentes de IA. Instálala en dos minutos.

$ pip install centinela

Works with zero backend — set CENTINELA_ENDPOINT=stdout and watch events stream to your console.

Three things, on by default

Evidence from the first run.

No agent is a black box once Centinela is recording. The SDK is fail-open by design: if the backend is slow or down, your agent never blocks, slows, or crashes.

REC

Record

Wrap your agent or drop in c.trace(...). Every action becomes a structured event: type, name, input, output, timing, status.

RPL

Replay

When an agent fails expensively, the context is usually already gone. Centinela keeps an auditable record you can replay, prove, and debug.

RDC

Redact

Turn on field-level redaction and sensitive payloads never leave your process — only structure, timing, and status ship. Compliance-safe by construction.

Real-time blocking · on the roadmap
Quickstart

Two minutes to your first trace.

  1. Install. One runtime dependency (httpx). Python 3.9+.
  2. Wrap, or instrument by hand. c.wrap() patches OpenAI, Anthropic and LangChain. Anything else uses the framework-agnostic trace API.
  3. See it immediately. No key yet? Point the endpoint at stdout and every event prints to your console.
from centinela import Centinela

c = Centinela(api_key="ctl_...", project="mi-agente")

# Option A — automatic wrapper
agent = c.wrap(agent)

# Option B — works with any framework
with c.trace("procesar_pedido") as t:
    t.log_action(type="tool_call", name="enviar_email",
                 blocked=True, human_review="approved")
    t.log_action(type="llm_call", model="claude-sonnet-4",
                 tokens=1234)
The plus

Your agents, already audit-ready.

The same trace that shows you what happened also doubles as your compliance layer. Centinela maps each recorded agent to a library of controls aligned with the frameworks your enterprise customers are starting to ask about.

ISO/IEC 42001NIST AI RMFEU AI Act

Build agents for clients? Ship the evidence with the product. Hand a customer a bilingual audit report and a Centinela Verified badge for their agent — a differentiator you can sell, not a fire drill you survive.

Centinela generates evidence from your real traces; it doesn't certify your company. 8 of 10 controls are measurable today from SDK signals. One transparency deadline is already close: under the EU AI Act, disclosure obligations for deployers apply from 2 Aug 2026, and provider watermarking from 2 Dec 2026 — the kind of thing it's cheaper to instrument now than to retrofit later.

Get started

Start recording today.

The SDK is open source and works on your machine in two minutes — no signup, no backend required. Hosted dashboards and the Verified seal are in early access.

View on GitHub