rea
Installation
SKILL.md
rea: binary and application reverse engineering
The rea MCP server runs on the host machine, not inside this container. Its tools are named mcp__rea__<tool>. After this skill loads, their names appear as deferred tools: fetch each tool's schema with SearchExtraTools (for example select:mcp__rea__open_binary, or keywords like rea strings), then invoke it with ExecuteExtraTool. Use the tool that directly answers the question; do not enumerate the whole catalog.
1. File paths: container view vs host view
rea opens files on the HOST filesystem, not inside the container. Only files under the bind-mounted data directory are visible to both sides; the exact host prefix is deployment-specific (check the compose file, or ask the operator via task chat when unknown).
Rules:
- Save or download every target file inside your run directory, with a relative path such as
rea/(your working directory is the run directory, and it sits under the bind-mounted data directory, so both sides can see it). Do not save to an absolute path outside the run directory: the Write tool refuses it. Files elsewhere in the container are invisible to rea. - Always pass the host-side path in rea tool parameters: take the absolute container path of that directory (
pwdin the run directory, plusrea/) and replace the container data directory prefix with the host prefix of the same mount. If open_binary reports the path as missing, you passed a container path: retry with the host-side equivalent of the same mount. - Verify the file exists on your side (ls) before opening it with rea.
2. Session lifecycle
open_binarywithpath(host path). For native binaries (ELF, PE, Mach-O) setprovider_id: "ghidra"; if the error lists candidate provider ids, retry with one of them. Packages and bundles (APK, IPA, HAP, zip, JAR, wasm, JavaScript) need no provider.- Run analysis tools (
search_strings,list_strings,search_procedures,inspect_native_*,project_managed_application_graph, family-specific graph tools, ...). Results are Evidence objects: they carry an evidence id, a normalized result, and a limitations section. Record anything task-relevant as facts in the knowledge graph and cite the evidence id in findings. close_binarywhen done. One open target per session; close before opening another. Never leave a binary open at the end of a run.