AI agents send emails, create records, and transfer funds with no isolation, no rollback, and no authorization gate. A single hallucinated DELETE reaches production the same way a SELECT does. UndoLog fixes that.
Classify each tool call by reversibility. Log everything to a persistent effect journal. Pre-register compensations before execution. Gate irreversible operations behind human approval. Treat every tool call like a database transaction.
No framework patches. No monkey-patching. The tool function never knows it is being managed.
Designed for LangGraph, CrewAI, OpenAI Agents SDK, or a custom agent loop. The unopinionated safety layer your agent infrastructure is missing.
In April 2023, a Replit user asked an AI agent to "delete the test database." The agent complied. The production database was gone.
In February 2024, a Google D: drive wipe incident affected over 200,000 users when a single agent action cascaded without checkpoints.
These are not stories about bad models. They are stories about missing infrastructure.
pip install undolog-sdkWalk through the installation guide to set up the proxy and engine. Then annotate your tools with @undolog_tool and UndoLog handles the rest.
The full docs follow the Diátaxis framework: tutorials, how-to guides, reference documentation, and explanation of the safety model. Start with the quick start or dive straight into the architecture decisions.
UndoLog is Apache 2.0 licensed. Contributions are welcome. Read the contributing guide to get started. Found a bug or have an idea? Open an issue.
If you discover a security vulnerability, email security@undobase.com. All reports are promptly addressed.