ZizkaDB 发布：用审计而非观测来调试 AI 智能体

Operational intelligence for AI agents

Dont Observe - Audit your AI Agent

When production agents drift, guesswork costs hours. ZizkaDB gives you replay, lineage, and drift alerts so you fix behavior before users notice.

Works with your stack

Python SDK

pip · zizkadb-sdk 0.2.8

TypeScript SDK

npm · zizkadb-sdk 0.2.7

Cursor MCP

zizkadb-mcp 0.1.7

LangChain

zizkadb-langchain 0.1.3

CrewAI

zizkadb-crewai 0.1.3

LiveKit

zizkadb-livekit 0.1.0

REST API

Any language

Open source

AGPL · self-host

Managed cloud

db.zizka.ai

Two ways to run ZizkaDB

Same product. Your infrastructure or ours.

Explore the dashboard your team gets — self-host the OSS stack or use managed cloud with Fleet.

Open source

Docker Compose on your infra. Activity, Behavior, Reports, and Suggestions out of the box.

Self-hosted · demo

Recent sessions

14:02:11tool_callsearch_policy_docsok

14:02:09llm_responseRefund window is 14 days…warn

14:01:58session_startcustomer-4821ok

Managed cloud

Hosted on db.zizka.ai — includes Fleet ranking across agents and projects.

Managed cloud · demo

Recent sessions

14:02:11tool_callsearch_policy_docsok

14:02:09llm_responseRefund window is 14 days…warn

14:01:58session_startcustomer-4821ok

Why ZizkaDB

Stop debugging agents in the dark.

Production agents need operational data — not just logs. ZizkaDB monitors behavior so your team can correct it.

How agents fail in prod

·Prompt tweaks silently change decisions
·Tool calls skip policy without anyone noticing
·Same question, different answers across sessions
·Logs show what happened — not why

What that costs you

·Wrong answers reach customers
·Engineering burns time on blind debugging
·Token spend climbs as agents retry
·Trust erodes before you find root cause

With ZizkaDB

✓Replay any session end-to-end
✓Measure drift against your baseline
✓Trace decisions with causal lineage
✓Catch regressions before they ship

See what changes when you can replay any session

Same incident. Without ZizkaDB you guess. With it you see the exact decision trail.

WITHOUT ZIZKADB

Your team today

Customer

Your bot told me refunds take 30 days. That is wrong.

Your team

Can you send a screenshot? We will check the logs.

Engineering

Logs show nothing useful. Maybe a prompt issue?

Hours lost. Still no root cause.

WITH ZIZKADB

Same incident, 2 minutes later

Drift alert

Refund answers changed after prompt v2 deploy.

Replay

Agent skipped policy doc, used outdated FAQ chunk.

Fix

Roll back prompt. Confirm baseline restored.

Root cause found. Prompt rolled back.

For teams building with agents

Integrate in an afternoon. Debug in minutes.

Log from Python, TypeScript, LiveKit voice agents, MCP, or REST. One API key per agent — no heavy infra project.

⚙️

Connect your stack

Python SDK, TypeScript SDK, LangChain, CrewAI, LiveKit voice, MCP, or REST. Self-host with Docker or use managed cloud.

🔍

Debug production incidents

Prompt change broke behavior? Replay the full session, trace the decision chain, find root cause.

🏢

Enterprise & private cloud

Single-tenant VPC deployment with Fleet dashboard, workshops, and dedicated support.

Trust and security

Production-ready from day one

📦

Open source

AGPL license. Inspect, fork, self-host free.

🏠

Self-host free

Docker Compose on your own infrastructure.

🔒

Checksum-backed events

Tamper-evident decision history.

🇪🇺

Managed cloud

db.zizka.ai when you want zero ops.

Pricing

Self-host free. Scale on cloud when you need to.

Full monitoring and session replay on every plan.

✓1 API key
✓Your infrastructure
✓Docker Compose
✓Community support

POPULAR

✓50k events / month
✓3 active API keys
✓Email support

✓100k events / month
✓10 active API keys
✓Priority support

✓Single-tenant VPC deployment
✓Up to 50 agents
✓Fleet dashboard and ranking
✓Install + integration workshop

Compare

Agent observability vs tracing platforms

Laminar, LangSmith, Langfuse, and Phoenix trace spans. ZizkaDB stores causal lineage, time travel over logged state, and native drift baselines.

Capability	Laminar	LangSmith	Langfuse	Arize Phoenix	ZizkaDB
Causal lineage (why()-style, not span tree)	~	~	✗	✗	✓
Time travel (state at T)	✗	✗	✗	✗	✓
Native drift/baseline	~ (rule-based)	✗	✗ (explicit gap)	✓ (strongest)	✓
Free self-hosting	✓ (Apache 2.0)	✗ (Enterprise only)	✓ (MIT)	✓ (ELv2, free but not OSI)	✓ (AGPL)
EU AI Act mapping	✗	✓ (first-party)	~ (third-party writeups; EU region only)	✗	✓ (first-party + EU-incorporated)
Total	2	2	1–2	2	5/5

Causal lineage (why()-style, not span tree)

Laminar

LangSmith

Langfuse

✗

Arize Phoenix

✗

ZizkaDB

✓

Time travel (state at T)

Laminar

✗

LangSmith

✗

Langfuse

✗

Arize Phoenix

✗

ZizkaDB

✓

Native drift/baseline

Laminar

~ (rule-based)

LangSmith

✗

Langfuse

✗ (explicit gap)

Arize Phoenix

✓ (strongest)

ZizkaDB

✓

Free self-hosting

Laminar

✓ (Apache 2.0)

LangSmith

✗ (Enterprise only)

Langfuse

✓ (MIT)

Arize Phoenix

✓ (ELv2, free but not OSI)

ZizkaDB

✓ (AGPL)

EU AI Act mapping

Laminar

✗

LangSmith

✓ (first-party)

Langfuse

~ (third-party writeups; EU region only)

Arize Phoenix

✗

ZizkaDB

✓ (first-party + EU-incorporated)

Total

Laminar

LangSmith

Langfuse

1–2

Arize Phoenix

ZizkaDB

5/5

~ = partial support. Verify competitor docs before external debates.

Fix behavior before production breaks

Managed cloud on db.zizka.ai, or self-host the open source stack anytime.