I just released TraceMotive v0.1, an open-source, local-first tracing and debugging tool for AI agent execution.
I built it because I kept running into the same problem:
When an agent behaves incorrectly, the final error often doesn't tell you enough about how the execution actually got there.
TraceMotive records agent runs as traces and spans so you can inspect the execution step by step.
What v0.1 includes
- Trace and Span collection
- Parent/child Span hierarchy
- Execution Timeline
- Span Inspector
- SQLite-backed local Collector
- OpenAI Agents SDK integration
- Privacy-first content capture defaults
- Bounded local transport and failure isolation
Install
pip install tracemotive
GitHub:
https://github.com/doraemonfv-glitch/tracemotive
PyPI:
https://pypi.org/project/tracemotive/
TraceMotive's own tracing data stays local by default. Model-provider traffic still depends on how the agent itself is configured.
Why I built it
v0.1 is intentionally focused on observation and debugging rather than automatic diagnosis.
The longer-term direction is to move from:
“Where did the error happen?”
toward:
“Where did the execution first start going wrong?”
That is the idea behind the long-term goal:
“The causal debugger for AI agents.”
I'd especially appreciate feedback from people building agents:
What information do you usually wish you had when debugging an agent execution?
Top comments (1)
If you build AI agents, I’d especially love to hear what kinds of failures are hardest for you to debug.