Ghost Hunter isn't the first AI bug-reproduction tool, similar internal systems exist at companies like Metabase, who deliberately gate them behind human review to prevent prompt injection from public issues. What's different here: Ghost Hunter is fully open-source and self-hosted rather than a proprietary internal tool, works with any GitHub repo with no custom infra, and has had real security measures implemented based on community feedback, not hidden limitations, but ones actively addressed.
I built Ghost Hunter because I kept losing time on the most annoying part of maintaining a project, trying to reproduce a bug from a vague issue report before I could even start fixing it.
The idea was simple: what if reproducing the bug was as easy as commenting bot/reproduce? It parses the issue with an LLM, spins up an isolated Docker sandbox, actually runs the repro steps, and posts the crash logs back, all automatically.
This is my first solo open-source project, and the process taught me a lot. Early testing surfaced real security concerns from the community: prompt injection risk, network isolation and rather than brush past them, I dug in and fixed them. That back-and-forth genuinely made the project better, and it's part of why I wanted to launch here: I'm hoping for more of that kind of honest feedback.
It's still early (v1.0.0), and I'm sure there's more to find. Would love to hear what you think: bugs, ideas, or just "this is dumb because X" are all welcome.