Gemma 4 Powers Local Claude Code Agents

The interesting part here is not just “local Gemma runs,” but that the model is being coerced into a full agent loop with shell access through compatibility glue that was originally built for Anthropic. It’s a practical proof that local open-weight models are getting close to real agent work, but the hardware tradeoff is still brutal.

• –The post highlights a real bottleneck for local agents: Claude Code’s large system prompt and tool protocol make small context windows unusable.
• –The claimed fix path is very specific and useful: `llama.cpp` + `-ctk q8_0 -ctv q8_0` + aggressive layer offload to fit a 31B model into 16GB VRAM.
• –This is more of a systems-engineering hack than a turnkey workflow; the model can act autonomously, but latency becomes the tax for squeezing it onto consumer hardware.
• –The bigger signal is architectural: agentic coding is becoming a protocol problem as much as a model-quality problem.
• –If the setup really holds up, it points to a future where local models can slot into existing agent shells with minimal application changes.