Who it is forA team that changes models or prompts and needs to know when an agent quietly starts acting differently.
The problemAfter a model or prompt update, behaviour shifts in small ways that nobody notices until users complain.
The outcomeDrift alerts against past sessions, and the ability to ask your coding assistant why an agent did something, answered from the audit record.
What you need first
- Docker
- Python or TypeScript SDK for integration
Steps
- Log sessions over time with the ZizkaDB SDK so there is history to compare
- Call db.baseline(agent) to set drift detection against past sessions
- After a model or prompt change, check the drift result before and after
- Add the MCP server to Cursor or Claude: command uvx with args zizkadb-mcp and ZIZKADB_HOST set to http://localhost:8000
- Ask in your assistant, for example why did support-bot call lookup_order?
- Replay sessions where drift shows up to confirm the change
Prompt for your agent
Connect the ZizkaDB MCP server (uvx zizkadb-mcp, ZIZKADB_HOST=http://localhost:8000). Compare <agent>'s behaviour before and after <date of model change> using its baseline, and list the sessions that drifted most with one sentence each on what changed.
Caveat. The README positions it as support for EU AI Act Article 12 record-keeping, not a compliance guarantee.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for ZIZKA-AI-SL/ZizkaDB and "Catch drift in agent behaviour and ask about it over MCP"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer