Does Gas Town 'steal' usage from users' LLM credits to improve itself?
| Source: HN | Original article
A new investigation into the open‑source “GasTown” proxy has uncovered that the tool may be siphoning off users’ LLM credits without clear consent. According to the recently released gastown‑release.formula.toml and beads‑release.formula.toml files, a default local installation automatically scans open issues on the github.com/steveyegge/gastown/actions repository. Each scan triggers a call to the user’s subscribed LLM provider—OpenAI, Anthropic or other services—thereby consuming API quota that appears on the user’s billing statement but is not tied to any explicit request.
The behaviour was first flagged by developers who noticed unexplained credit depletion after installing GasTown. A deeper look at the configuration revealed that the issue‑review routine runs on a 20‑minute interval, a cadence echoed in unrelated discussions about Google Antigravity’s backend usage limits. Steve Klabnik’s recent blog post describes GasTown as “simultaneously boring and opaque,” hinting at the lack of transparency that now seems to have concrete cost implications.
Why this matters goes beyond a surprise bill. GasTown is marketed as a lightweight, locally hosted gateway for LLM experimentation, a niche that many Nordic startups and research labs rely on to stretch limited free‑tier credits. If the tool is silently expending those credits to “improve itself”—presumably by feeding usage data back to the maintainer’s own models—trust in community‑driven AI infrastructure erodes, and budgeting for AI projects becomes riskier.
The community response is already shaping the next steps. Stevey Yegge, the project’s primary maintainer, has opened a GitHub issue promising a patch that will make the issue‑scanning feature opt‑in rather than default. Watch for an updated release candidate within the next week, and for broader discussions on auditability standards for open‑source LLM proxies, which could influence how Nordic firms vet third‑party tooling in the months ahead.
Sources
Back to AIPULSEN