ToCode

ToCode exports a binary or IDA database into a source-like project tree: raw recovered C, matching assembly, function summaries, section data, optional IDA database, and metadata that coding agents can read directly.

Why

AI models are strong at coding, especially when they can traverse large codebases and accumulate context with subagents and other strategies. When we use these agents to assist with reverse engineering, we usually provide tools through MCP or other means so the coding agent can learn and build strategies around tools such as IDA and r2. This approach adds limitations and constraints to how the agent behaves, and it increases the need for deep, complex reasoning.

There should be a better way to improve this scenario so that even smaller models can perform well on this kind of work.

The idea behind ToCode is simple: use a disassembler such as IDA to create a source-code-like project for a given binary, with a pre-built AGENTS.md so most coding agents start with precomputed context. ToCode also produces rich .json files with important metadata.

With this approach, even tiny models can perform well without being connected to MCP-like tool calls, because ToCode provides exactly what coding agents are good at working with: code.

Export layout

The exported project contains the following structure:

sample_decompiler/
  AGENTS.md
  CLAUDE.md
  src/raw/**/*.c
  src/raw/**/*.asm
  src/raw/**/*.summary
  include/*.h
  include/*.types.h
  data/*.bin
  data/variables.json
  data/variables_interesting.json
  function-index.json
  functions.json
  types.json
  sections.json
  strings.json
  imports.json
  exports.json
  relocations.json
  reachable.json
  cluster-graph.json
  triage.json
  project.json
  export-manifest.json
  tocode.log

Path	Description
src/raw	Decompiled C-like output, assembly, and summaries. Grouped by call-graph cluster, or by the original source file/directory when the binary has debug info (DWARF).
include	Generated headers, including *.types.h with the structs/enums/typedefs recovered from the binary.
data	Raw section dumps and variable metadata.
types.json	Catalog of types recovered from the binary's debug info or type library, with C declarations.
*.json	Functions (with recovered types and original source decl file/line), sections, strings, imports, exports, relocations, reachability, clusters, triage, project metadata, and export manifest.
tocode.log	Export log with checkpoint, resume, and per-function render history.
AGENTS.md / CLAUDE.md	Instructions for agents analyzing the exported binary.
src/tree	Optional scanner-friendly C output when --tree is used.

Supported backends

Three backends are supported, selected with --backend (default auto, which prefers them in this order):

1. IDA – uses the ida-domain/idapro Python libraries (best decompilation quality).

2. radare2 – uses r2pipe with the r2ghidra decompiler.

3. angr – a pure-Python fallback with no external tooling, so ToCode still runs when neither IDA nor radare2 is available. It is an optional extra (it is large: ~450 MB of native dependencies), so it is not installed by default. Get it in any of these ways:

   bash ./install.sh --all                        # recommended: installs every backend
   pip install tocode-cli[angr]                   # or, with pip directly
   uv tool install --force --editable '.[angr]'   # or, from a local checkout

   The angr export is structurally identical to the others (same files and metadata); its pseudo-C is lower quality than Hex-Rays or r2ghidra.

4. Binary Ninja, is opt-in (--backend binja) and never chosen by auto, because it drives a running Binary Ninja instead of reading a file. See Binary Ninja below.

Other disassemblers may be added in the future.

Using

ToCode supports Windows, Linux, and macOS with Python 3.10 or newer.

On Windows PowerShell:

powershell -ExecutionPolicy Bypass -File .\install.ps1

On Linux or macOS:

bash ./install.sh

Add --all or --full (PowerShell: -Full) to also install the angr fallback backend, so ToCode works out-of-the-box even without IDA or radare2:

bash ./install.sh --all

Manual setup (requires uv):

git clone https://github.com/buzzer-re/ToCode
cd ToCode
uv sync --locked
uv tool install --force --editable .

Example

tocode firmwareX.bin -o firmwareX_decompiled/
cd firmwareX_decompiled/
codex 

# Inside your agent shell, type your goals, e.g.: "Give me a brief overview of the boot process of this firmware."

From an ongoing RE work

Interrupted exports save progress automatically. Rerun the same command to resume from cached function renders, or add `--restart` to ignore the saved checkpoint and start over.
tocode firmwareX.bin.i64 -o firmwareX_decompiled/
...

Binary Ninja

The Binary Ninja backend exports the binary you already have open in Binary Ninja, emitting its Pseudo C. It does not need a Binary Ninja headless license: ToCode connects to the running GUI over a small RPC bridge.

Since I don't have access to Binary Ninja headless, run this script once in Binary Ninja's Python console to open an RPC server ToCode can connect to:

import threading, rpyc, binaryninja
from rpyc.utils.server import ThreadedServer

class ToCodeService(rpyc.Service):
    exposed_binaryninja = binaryninja
    def exposed_bv(self):
        return bv  # the currently focused view
    def exposed_eval(self, cmd):
        return eval(cmd)

server = ThreadedServer(
    ToCodeService,
    hostname="127.0.0.1",
    port=18812,
    protocol_config={"allow_all_attrs": True},
)
threading.Thread(target=server.start, daemon=True).start()
print("ToCode RPC server listening on 127.0.0.1:18812")

It exposes the whole Binary Ninja Python VM with no authentication, so keep it bound to 127.0.0.1. The binja-headless plugin exposes the same interface if you prefer a packaged option.

Then drive it from ToCode:

tocode --backend binja --list-binja              # list open views with an index
tocode --backend binja -o out/                   # export the focused view
tocode --backend binja --binja-view 1 -o out/    # export a specific view by index
tocode --backend binja --all-views -o out/       # export every open view, one folder each

Use --binja-host / --binja-port (or TOCODE_BINJA_HOST / TOCODE_BINJA_PORT; default 127.0.0.1:18812) to reach Binary Ninja on another machine.

Already scripting inside Binary Ninja? Skip the server and export the live view directly:

import sys; sys.path.insert(0, "/path/to/ToCode/src")
from tocode import export_from_binaryview
export_from_binaryview(bv, "out/")

Development

This tool was built using agentic coding, so if you plan to help, I strongly advise doing the same.

Before changing ToCode, have Python, uv, ruff, mypy, pytest, and compileall available. For backend work, also have IDA or radare2 installed, depending on what you are touching.

The main instructions for agents are in AGENTS.md. Read it before starting, and make sure the local quality gate passes before proceeding.

Quality Gate

Run the local CI gate before opening a PR:

./ci-local.sh

On Windows PowerShell:

powershell -ExecutionPolicy Bypass -File .\ci-local.ps1