Const24.GhidraSharp 0.7.0

GhidraSharp

Drive headless Ghidra from C#: import a binary, analyse it, decompile, walk cross-references, and write findings back. A typed .NET client over a small gRPC bridge, with no Python in the chain.

There is no first-party C# binding to Ghidra. The community route is the Python ghidra_bridge, and Ghidra's own cross-language protocol, TraceRmi, is scoped to debugging rather than static analysis. This fills that gap with a small typed surface over the core reverse-engineering operations.

  ┌─────────────────────┐   gRPC (protobuf)  ┌─────────────────────────────┐
  │ Const24.GhidraSharp │   ──────────────>  │ GhidraSharpServer (Java)    │
  │ C# client (net8/10) │   <──────────────  │ Ghidra-as-library, headless │
  └─────────────────────┘                    └─────────────────────────────┘

Requirements

• .NET 8 or later. The client targets net8.0 and net10.0; the MCP tool is net10.0.
• JDK 21 or later, to run the server.
• Ghidra 12.1 or later, with GHIDRA_INSTALL_DIR pointing at it.

The JDK and Ghidra are not shipped with the packages.

Install

Client and server, with the server bundled into your app's output so dotnet publish carries it and StartAsync finds it:

dotnet add package Const24.GhidraSharp.Server

Client only, if you connect to a server you run yourself:

dotnet add package Const24.GhidraSharp

Quickstart

From a raw firmware dump to its decompiled functions. No server to start, no project to manage.

using Const24.GhidraSharp;

// StartAsync finds the bundled server, runs it, and stops it on dispose.
await using var server = await GhidraServer.StartAsync();
var ghidra = server.Client;

// Import and analyse. languageId is the target's processor — here a Renesas SH-2A.
await ghidra.OpenProgramAsync(@"C:\firmware\ecu.bin", languageId: "SuperH:BE:32:SH-2A");

var functions = await ghidra.ListFunctionsAsync();
foreach (var fn in functions)
    Console.WriteLine($"  {fn.EntryPoint}  {fn.Name}");

var dec = await ghidra.DecompileAtAsync(functions[0].EntryPoint);
Console.WriteLine(dec.CCode);

languageId is the one Ghidra-specific input: SuperH:BE:32:SH-2A, ARM:LE:32:v7, x86:LE:64:default. ListLanguagesAsync("SuperH") enumerates them if you are not sure.

That example is transient — nothing touches disk. CreateProjectAsync writes a Ghidra project you reopen in seconds and can save renames into. For a headerless image it also takes baseAddress and entryPoint, which matter as much as the processor: at the loader's default of 0, the image's own pointers resolve to nothing and analysis finds almost no code.

What is bridged

The surface grows one RPC at a time, as consumers need it. Every client method is the name below plus Async.

Program

• Ping — liveness, Ghidra version, server version
• OpenProgram — open an analysed project, or import a binary transiently
• CreateProject — import into a persistent project (.gpr/.rep), analysed and saved; takes baseAddress/entryPoint for a headerless image
• SaveProgram / CloseProgram — persist edits; release the on-disk lock
• ListLanguages — the processor languages Ghidra supports, i.e. where a languageId comes from
• ListMemoryBlocks — the program's sections: name, range, size, permissions

Code

• DecompileFunction / DecompileFunctions — function → C; the second streams a whole program
• ListFunctions — every function and its callees, built to query client-side with LINQ
• GetFunction — one function in full: typed signature, parameters, locals, callers
• GetInstructions — the disassembly listing: mnemonic, operands, bytes
• GetInstructionDetail — one instruction's structured operands and raw PCode
• ReadBytes — raw program memory, for a pure-C# byte/table layer

Symbols, references, data

• ListSymbols / GetSymbolsAt — symbols by name or address
• GetReferencesTo / GetReferencesFrom — cross-references
• GetFunctionReferences — every reference out of a function's body, its table data-refs included
• ListImports / ListExports — the import table grouped by library, and the exports; each import carries the EXTERNAL-space address that GetReferencesTo resolves to its call sites
• FindStrings — defined strings whose decoded text matches, each with its xrefs — the concept → code loop
• GetDataAt / ListDataTypes / ApplyDataType — defined data and data types
• GetDataType — a named struct/union/enum's full member layout

Writing findings back — needs writable: true, in memory until saved, and each returns the value it replaced

• RenameSymbol — name a function, label or global
• SetComment / GetComments — comments, all five Ghidra types
• SetBookmark / GetBookmarks — bookmarks

The escape hatch

• RunScript — run any GhidraScript and capture its output

Architecture-agnostic by construction: it forwards a Ghidra language id, so the same code drives any processor Ghidra supports. Every result is a hand-written, documented record — the generated gRPC types stay internal, and the XML docs explain each Ghidra concept to a .NET developer who has never opened Ghidra. A failed call throws GhidraException carrying Ghidra's own diagnosis; a cancelled one throws OperationCanceledException, as a .NET caller expects.

Not bridged, on purpose: the live Ghidra object graph, the decompiler's internals beyond C text and raw PCode (no HighFunction), and in-process scripting semantics.

Going further

Links are absolute so they work from nuget.org as well as from the repository.

For agents

A stdio MCP server that exposes the same surface as 30 tools an agent calls directly. It keeps several programs open at once, caches analysed programs by content hash, and journals every rename and comment outside that cache so the analysis survives a cache wipe. run_script is off by default and not even listed when off.

dotnet tool install -g Const24.GhidraSharp.Mcp

{ "mcpServers": { "ghidra": {
    "command": "ghidrasharp-mcp",
    "args": ["--server-dir", "C:\\path\\to\\ghidrasharp-server"]
} } }

The rest — flags, headerless images, what each tool is for — is in docs/mcp/using.md.

Correctness

GhidraSharp and pyghidra drive the same Ghidra with the same calls, so for any operation the two must agree byte for byte; anything else is a bridge bug. bench/ is that comparison, over eight capabilities and six instruction sets, on targets it builds from tools you already have — no ROMs, no private data, one command:

python bench/verify.py

The latest run is in bench/REPORT.md.

Security

The server is unauthenticated and exposes RunScript — arbitrary GhidraScript — plus file and memory access. It binds to loopback only. It is a local tool driven by a client on the same host. Never expose its port. The rest of the posture, and how to report something, is in SECURITY.md.

Build and test

dotnet build GhidraSharp.slnx
dotnet test tests/GhidraSharp.Tests -- --filter-not-trait "Category=Integration"
cd server && ./gradlew test

The integration tier needs a real Ghidra and is gated; it skips silently without one. CLAUDE.md is the maintainer's guide and covers that and the rest of the build.

License

Apache-2.0, the same as Ghidra, which the server links against.