Project
Local LLM / OpenClaw Experiments
Experiments with local models, tool calling, agent orchestration, and constrained hardware.
Problem
Local LLMs are increasingly capable, but they are not drop-in replacements for hosted frontier models. The product and workflow need to fit the hardware.
Why I Built It
I wanted hands-on evidence about where local models are useful for agents, tool calling, and personal productivity workflows.
Architecture
The experiments compare local model runtimes, tool-call contracts, prompt structure, and agent orchestration boundaries on constrained hardware.
Tech Stack
- OpenClaw
- LM Studio
- Local GGUF models
- Ubuntu
- Tool calling
Current Status
Research and experimentation.
Experiment Notes
The most useful comparison is not raw model score. It is whether a local model can complete a constrained workflow with acceptable latency, predictable tool use, and a failure mode that is easy to inspect.
Each experiment tracks the model, prompt structure, available tools, hardware constraints, task shape, and the point where human review becomes necessary.
What I Learned
Small local models work best when they are asked to perform narrow, structured tasks with explicit tools.
Next Steps
Document hardware limits, build a small repeatable task benchmark, and test one useful local workflow end to end.