# ZizkaDB

> Operational audit database for AI agents with causal lineage, session replay, drift detection, and EU AI Act compliance. Self-host free (AGPL) or use managed cloud.

ZizkaDB is an operational audit database for AI agents built by ZIZKA AI S.L., a company incorporated in Málaga, Spain. It stores agent decisions, tool calls, and outcomes with explicit causal links, enabling teams to replay sessions, trace root causes, and detect behavioral drift in production. The core stack is open source under AGPL-3.0 and can be self-hosted with a single Docker Compose command, while a managed cloud option at db.zizka.ai and private VPC enterprise deployments are also available.

## What It Is

ZizkaDB is an agent observability and auditing platform that sits alongside — not instead of — vector databases and distributed tracing tools. Where span-based tracers like LangSmith or Langfuse record what happened, ZizkaDB records *why* it happened: each logged event carries a `parent_id` that links it causally to the step that triggered it. The result is a queryable decision graph rather than a flat log, enabling `db.why(event_id)` to walk backward from any agent output to the originating user message, wrong tool call, or stale context chunk. The platform also supports time-travel queries (`db.at(timestamp)`) to reconstruct exactly what an agent knew at a given moment, semantic search over agent history, behavioral baseline comparison, and GDPR erasure via `db.forget()`.

## Core Capabilities

- **Causal lineage (`why()`)** — walk any decision chain back to root cause via explicit `parent_id` links stored in Postgres
- **Time travel (`at()`)** — reconstruct agent state at any past timestamp
- **Drift detection (`baseline()`)** — compare recent sessions to each agent's behavioral baseline; verdicts range from Stable to Significant
- **Session replay** — end-to-end replay of any production session from the dashboard
- **Semantic search** — search over agent history using embeddings (OpenAI model, configurable)
- **GDPR erasure** — `db.forget()` deletes events by metadata filter on request
- **Tamper-evident logs** — per-event SHA-256 checksums for audit integrity
- **EU AI Act Article 12 mapping** — first-party record-keeping controls, EU-incorporated operator

## Integration and Setup Path

ZizkaDB is designed to integrate in an afternoon. The self-hosted stack starts with one command (`bash scripts/quickstart.sh`) and opens a dashboard at `localhost:3001` with no email signup required. Managed cloud requires only an email OTP and an API key. SDKs are stateless and accept `agent`, `session_id`, and `event_id` as explicit parameters.

Supported integration paths include:
- **Python SDK** (`zizkadb-sdk` 0.2.8 on PyPI)
- **TypeScript SDK** (`zizkadb-sdk` 0.2.7 on npm)
- **LangChain adapter** (`zizkadb-langchain` 0.1.3)
- **CrewAI adapter** (`zizkadb-crewai` 0.1.3)
- **LiveKit voice agents** (`zizkadb-livekit` 0.1.0) — one LiveKit call becomes one Session in the Activity view
- **MCP server** (`zizkadb-mcp` 0.1.7) — works with Cursor and Claude Desktop
- **REST API** — any language via HTTP

## Deployment Model

ZizkaDB runs in three modes using the same open-core engine:

1. **Self-hosted (AGPL)** — Docker Compose on any infrastructure, free forever, community support, one API key
2. **Managed cloud** — hosted at db.zizka.ai, includes Activity, Behavior, Reports, Suggestions, and Fleet dashboards; event quotas and multiple API keys per plan
3. **Enterprise VPC** — single-tenant deployment in the customer's own cloud region, fleet dashboard and ranking, audit export with checksums, commercial license, and a supported 28-day install path (Day 0 discovery → Day 7 staging → Day 14 integration workshop → Day 28 handoff with Zizka access revoked)

The enterprise page notes that SSO, SAML, and automated alerting are on the roadmap; pilots use dashboard OTP plus VPN access.

## Update: v0.2.8

The latest GitHub release is **v0.2.8**, published 2 September 2026, described as "New Production version of All SDKs except LiveKit." Current SDK versions are: Python SDK 0.2.8, TypeScript SDK 0.2.7, MCP 0.1.7, LangChain/CrewAI 0.1.3, and LiveKit 0.1.0. The repository was created in May 2026 and last pushed in September 2026, indicating active early-stage development. The GitHub repository has 114 stars and 22 forks as of the last recorded update.

## Features
- Causal lineage with db.why() — walk any decision chain back to root cause
- Session replay — end-to-end replay of any production agent session
- Behavioral drift detection with per-agent baselines
- Time-travel queries with db.at() to reconstruct agent state at any timestamp
- Semantic search over agent history
- GDPR erasure via db.forget() by metadata filter
- Tamper-evident per-event SHA-256 checksums
- EU AI Act Article 12 record-keeping mapping
- Fleet dashboard for multi-agent visibility
- Self-host free with Docker Compose (AGPL)
- Managed cloud at db.zizka.ai
- Enterprise single-tenant VPC deployment
- Python SDK, TypeScript SDK, LangChain, CrewAI, LiveKit, MCP, and REST integrations
- MCP server for Cursor and Claude Desktop
- Audit export with checksums (CSV/JSON)
- No framework lock-in

## Integrations
Python, TypeScript, LangChain, CrewAI, LiveKit, Cursor, Claude Desktop, REST API, Docker, OpenAI (embeddings), PostgreSQL

## Platforms
LINUX, WEB, API, DEVELOPER_SDK, CLI

## Pricing
Open Source, Free tier available

## Version
v0.2.8

## Links
- Website: https://db.zizka.ai
- Documentation: https://db.zizka.ai/docs
- Repository: https://github.com/ZIZKA-AI-SL/ZizkaDB
- EveryDev.ai: https://www.everydev.ai/tools/zizkadb