Local Models Triage the OpenClaw Repo for Free
Deploy Gemma‑4‑26b‑a4b or Qwen3.6‑35b‑a3b locally on a DGX Spark to replace paid APIs for OpenClaw issue triage.
Set up localpager‑agent on your server, configure reposhell, and point it to your OpenClaw repo for real‑time issue classification.
Summary
On June 22 2026, Onur Solmaz described how he built a real‑time notification system for the OpenClaw repo using local Gemma and Qwen models in an agent harness. The system runs on a DGX Spark with 128 GB unified memory (NVIDIA GB10) and can serve gemma‑4‑26b‑a4b at hundreds of tokens per second, replacing the $200 / month ChatGPT Pro plan by batching issue and PR classification locally. The agent harness, called localpager‑agent, uses a restricted reposhell to allow read‑only GitHub‑style commands, preventing malicious model behavior. The pipeline accepts a Hugging Face model ID and runs server‑side evaluation on held‑out data, supporting Whisper, IBM Granite, Cohere Transcribe, Wav2Vec2, HuBERT, SpeechBrain, and custom evaluators. The system classifies issues into finite categories like local_models, self_hosted_inference, acp, agent_runtime, codex, ui_tui, and uses the agent to output a JSON schema. The architecture includes a SQLite queue, a FastAPI route, and Discord notifications filtered by user policy. Future work includes expanding to multi‑model inference and improving classification accuracy. The approach demonstrates that local open‑weight models can replace paid APIs for triage tasks.
Key changes
- Deployment of local Gemma‑4‑26b‑a4b and Qwen3.6‑35b‑a3b for issue triage on OpenClaw.
- Use of DGX Spark with 128 GB unified memory (NVIDIA GB10) to achieve hundreds of tokens per second.
- Replacement of $200/mo ChatGPT Pro plan with local inference, batching every 2–6 hours.
- Implementation of localpager‑agent harness with reposhell for read‑only GitHub commands.
- Classification pipeline supports Whisper, IBM Granite, Cohere Transcribe, Wav2Vec2, HuBERT, SpeechBrain, and custom evaluators.
- SQLite queue and FastAPI route manage job scheduling and Discord notifications.
- JSON schema output with success flag and data payload for each classification.
- Demonstrated reduction of API costs and near‑instantaneous notifications for triage.