Malvex

Malvex is a Rust-based, local-first orchestration harness published on GitHub by JoasASantos. It connects AI agents to reverse-engineering and malware-analysis tools through the Model Context Protocol (MCP). Given a binary, memory dump, document, or packet capture, it triages the sample, runs parallel agents over backends such as IDA, Ghidra, radare2, Frida, volatility3 and YARA, and returns IOCs, MITRE ATT&CK mappings, detection rules, and attack-chain graphs.

What It Is

Malvex is a malware-analysis framework that drives an AI agent runtime (Claude Code by default, plus Codex CLI, Hermes, Ollama, or a custom command) to operate existing reverse-engineering tools. The harness runs analysis locally. Data transmission still depends on the selected agent runtime, model provider and optional external integrations; local orchestration is not a universal no-upload guarantee. Targets include native binaries (PE, ELF, Mach-O), packed or obfuscated samples, memory dumps, Office/PDF/RTF documents, pcaps, APK/DEX files, firmware, shellcode, .NET assemblies, and scripts.

Pipeline and Evidence Model

Runs move through triage, static, dynamic, network, intelligence, and validate stages, each spawning parallel agents that share state via MCP. The project's stated rule is that LLM output is not evidence: model hypotheses stay analyst leads until tied to static or runtime evidence, and a validator agent rejects unsupported conclusions. Each session writes target.json, findings.json, a summary, attack-graph files in mermaid and JSON, logs, and optionally IOCs and generated YARA/Sigma rules.

MCP Server Mode and Setup

The CLI offers doctor, mcp-setup, run, extract, graph, sessions, reparse, coverage, and repl commands. Running serve exposes analysis tools over MCP (JSON-RPC over stdio) so agents like Claude Code, Cursor, Windsurf, Cline, or Codex can call tools such as triage, decompile, YARA scan, IOC extraction, and Volatility3 listing. Setup scripts exist for macOS/Linux, Windows, and Kali; commercial tools like IDA and Binary Ninja are detected but never auto-installed. Dynamic analysis runs natively only when the host OS matches the target; otherwise agents emit a plan for the required runtime.

License and costs

The README states MIT licensing. Local harness access is separate from agent/model usage, commercial reverse-engineering tools, threat-intelligence services and lab compute. Offline operation requires a local model and suitable local tools; choosing cloud agents or external lookups changes that boundary. Android files are analysis targets, not evidence of an Android app distribution.