REA and OpenCode: local reverse engineering setup

This guide connects REA and Ghidra to OpenCode through MCP, with Strata serving Qwen 3.8 Flash Next 125B as a local LLM. The presenter reports running the workflow offline with no cloud charges. Setup covers Node 24, Ghidra 12.1.4 and Java 21. Most hardware references identify an RTX 5060 Ti with 16GB VRAM, although the introduction says RTX 4060 Ti.

The 3-bit mixture-of-experts setup keeps about 47GB of experts in system RAM and caches about 8.8GB on the GPU. The presenter reports that Strata's tuning raises generation speed from 39 to 77.7 tokens per second on this machine. Two configuration fixes address repeated context compaction: reducing REA's enabled tools from 170 to 19, then increasing context from 64,000 to 128,000 tokens while capping thinking at 8,000 tokens per turn.

Four reconstruction tests show how the AI agent inspects binaries and compares rebuilt behavior with the originals. The presenter reports matching calculator outputs, compatible scrambled notes files and a terminal game's hidden cheat code. The calculator retains function names; the other three binaries are stripped. Kilo, the final text editor test, remains incomplete after about 85 minutes, with cursor and search differences. Manual checks also leave search, the save message and multiline comments untested. These results describe the demonstrated apps rather than establish reliability for arbitrary software.