Onur, an OpenClaw maintainer, had a problem: hundreds of issues and PRs land daily, and he needed to spot P0 fires in his domain instantly. Paying for ChatGPT Pro to poll every few hours felt wasteful and slow. So he asked a different question—could the NVIDIA GB10 sitting on his desk, with 128GB of unified memory, do the job for the cost of electricity?
The answer turned out to be yes, but the approach is what makes it interesting. This isn’t 2023-style classification via a single API call with a JSON schema. Onur built an agent harness using Pi, where local models like Gemma-4-26b and Qwen3.6-35b actively investigate each issue. The agent doesn’t just read the title and body; it can use a restricted shell called reposhell to grep through the codebase, check git history, and read relevant files before committing to a label.
That restricted shell is the clever safety move. Giving a raw bash prompt to a model processing untrusted, potentially prompt-injected GitHub issues is asking for trouble. Reposhell only permits read-only commands like ls, cat, grep, and git show, and it tells the model it’s using bash so existing tool-use patterns work without modification. In one saved session, Qwen3.6-35b correctly classified a tricky PR about Kimi tool-call rewriting only after it explored the relevant source files to confirm the stop reason handling logic.
The whole setup runs faster than batch processing with a paid API and, critically, can’t be taken away. With the sting of Anthropic removing Claude Fable 5 still fresh, Onur’s experiment makes a pointed argument: if your business depends on AI classification pipelines, owning the stack and running capable local models is moving from a hobbyist’s dream to a legitimate infrastructure choice.
