Corvus

Corvus

Reverse engineering for AI agents. Corvus is an MCP server that lets your agent read, trace, compare and patch compiled programs: .NET assemblies and native PE, ELF and Mach-O binaries. One self-contained binary, no runtime to install, nothing uploaded; by default nothing it reads is executed.

Note

Corvus is under active development. Tools, parameters and output can change between releases; breaking changes are listed in the release notes. Bug reports and feedback are welcome in issues. If Corvus is useful to you, a star helps other people find it.

Quick start · How it works · What you can analyze · Showcases · Why Corvus · Documentation

Quick start

Windows (PowerShell):

irm https://raw.githubusercontent.com/Albasha0x0/Corvus/main/install.ps1 | iex

Linux / macOS:

curl -fsSL https://raw.githubusercontent.com/Albasha0x0/Corvus/main/install.sh | sh

The installer verifies the download against the release's SHA256SUMS and registers Corvus with the MCP clients it finds. As a plugin (server and skill; needs the .NET 10 SDK): /plugin install corvus --marketplace Albasha0x0/Corvus in Claude Code, codex plugin marketplace add Albasha0x0/Corvus then corvus in /plugins in Codex (installation). Restart your client and ask:

Load app.dll with Corvus and explain how it builds its export file. Cite the methods you read.

From a terminal: corvus decompile_member --load app.dll member=Ns.Type::Method (command line).

How it works

Your agent calls Corvus over MCP. Corvus answers with bounded, structured results: every list pages, every error names the next step, and long results are kept for read_result. The agent follows calls and references, explains the behaviour, and can patch a copy and save it.

Native code is analyzed by Corvus's own engine (x86, x64, ARM32, ARM64), with IDA or Ghidra as optional decompilers when installed. Their databases and output are kept, so the second question about a binary is fast.

What you can analyze

Target	What you get	Needs
.NET assemblies	C# decompilation, IL, references, call graphs, build comparison, C# and IL patching, saving	nothing
Unity and Il2Cpp games	metadata, addresses, string literals, IDA / Ghidra scripts, C++ bindings	nothing
Native binaries (PE, ELF, Mach-O)	functions, disassembly, CFG, cross-references, values, RTTI, PDB types, similar functions	nothing
Native pseudocode	decompilation, engine types, frames, scripting	IDA 9 (idalib) or Ghidra 12
NativeAOT and ReadyToRun	method names, MethodTables, strings, precompiled code	nothing
Native patching	names, comments, byte and assembly patches, patched copies, carried to new builds	nothing
Containers	single-file bundles, NuGet, APK, ASAR, Unity WebGL, HAR captures, installers (MSI, NSIS, Inno Setup, WiX Burn), frozen Python programs, firmware images; CycloneDX SBOM, YARA-X matches	nothing
Android packages	manifest, resources, DEX, smali, Java (JADX), JNI bindings, devices over adb	JADX for Java, adb for devices
Crash dumps	Windows minidumps, Linux and macOS core dumps, .NET exceptions, stacks, heap; crash text	nothing
Running programs	sessions with stop points, call recording, code coverage, activity traces, snapshots and child processes on Windows, Linux and macOS	Frida 17.24.3 or newer 17.x; off by default

Full list: features and the tool reference (165 tools, 15 prompts).

Showcases

• Decompile, rebuild, run: an assembly exported as a C# project that builds with warnings as errors, runs with the same output, and is measured against the original.
• Inside a single-file app: a .NET single-file executable extracted without running it, its assemblies extracted byte for byte.
• Where SQLite makes its random numbers: a crypto constant in a native SQLite build traced to sqlite3_randomness and its callers, recorded as an evidence bundle and re-verified.
• A Json.NET error traced to its code: "Self referencing loop detected" followed from the message to the deciding setting and every caller, with the evidence re-verified.

Each step in the showcases and guides is checked by a test.

Why Corvus

• Works with nothing installed. Native analysis, .NET decompilation, containers and dumps are built in.
• Fast. Results come from saved indexes and engine databases; warm calls answer in milliseconds.
• Sized for agents. Every result fits a token budget per part, with paging and kept results.
• Edits, not just reads. Patch .NET and native code in memory, review the diff, save a verified copy, and carry names and patches to the next build.
• Every platform. Windows, Linux (glibc and musl) and macOS, x64 and arm64: win-x64, win-arm64, linux-x64, linux-arm64, linux-musl-x64, linux-musl-arm64, osx-x64 and osx-arm64.

Supported clients: Claude Code, Claude Desktop, Cursor, Windsurf, VS Code, VS Code Insiders, VSCodium, Cline, Roo Code, Kilo Code, Visual Studio, Codex, Gemini CLI, Qwen Code, GitHub Copilot CLI, OpenCode, Crush, Zed, LM Studio, Amazon Q, Kiro, Antigravity, Devin, Command Code, Grok Build, OMP and Pi. Any other client: corvus config.

Documentation

Installation · Configuration · Features · Tool reference · Command line · Errors · Troubleshooting · FAQ · Release notes · All pages

Guides: first session · decode and patch · new build.

Responsible use

Analyze software you own or are permitted to analyze. Decompiling, patching or redistributing third-party code can be restricted by its license and by law where you live.

Contributing · Security · MIT license

Star history