Reddit thread maps the best tools for running coding agents on 8–12GB local models
salgado18 · reddit · 2026-07-22
The author asks what tools and harnesses people use to run complex coding tasks on small local models that fit in 8–12 GB of VRAM.
They argue that, on constrained models, the harness matters as much as the model itself: persistent memory, code search, call graphs, AST-based navigation, and fewer tool calls can reduce token waste and improve effectiveness.
Tools mentioned include:
- ai-memory for centralized wiki-style memory
- sigmap for code knowledge and search
- Playwright CLI for webpage navigation without screenshots
- CachyLLama, a llama.cpp fork with aggressive caching for AMD APUs
More from coding & agent
- A single Markdown rules file can fix Claude Code’s worst habits — eyishazyer · 2026-07-22
- Prompt library idea: delete rules, test variants, keep only what matters — _ScottCondron · 2026-07-22
- Anthropic adds Teach Claude a skill to turn screen recordings into reusable tasks — TawohAwa · 2026-07-22
- Revid says a watched GitHub repo can now auto-generate a full update video — tibo_maker · 2026-07-22
- LLM-as-a-Verifier uses continuous scores to improve agent evaluation — solyarisoftware · 2026-07-22
- Open-source `CLAUDE.md` file turns Karpathy’s coding advice into guardrails for Claude Code — socialwithaayan · 2026-07-22