Tested on Ubuntu 24.04.

To install the Python and Rust toolchain dependencies, run make install and follow the prompts if requested.

To translate a single C project to a Rust workspace, ensure it uses Cmake as a build system, place it in the examples folder and run:

make examples/C-project-name/translate \
  OPENROUTER_API_KEY="your-key" \
  MODEL="openai/gpt-5.1" \
  TRANSLATION_TEST=smoke

To run translation on a C project folder, it must be placed in the top-level examples folder.
The toolkit assumes the official DARPA TRACTOR folder structure for the C projects that will be translated:

📦IDEAS
 ┣ 📂src/ideas  # Core library
 ┗ 📂examples   # Project folders go here
   ┣ 📂C-project-name # A single C project folder with an arbitrary name
   ┃ ┣ 📂test_case
   ┃ ┃ ┣ 📂include
   ┃ ┃ ┣ 📂src
   ┃ ┃ ┗ 📄CMakeLists.txt # Must be correct and complete
   ┃ ┗ 📂test_vectors
   ┃   ┣ 📄some-name.json
   ┃   ┗ 📄other-name.json
   ┗ 📂other-C-project-name

This follows the standard evaluation convention in the DARPA TRACTOR project, which can be consulted in more detail here (including many C projects in examples).

The translation tool identifies each Cmake target (library or binary), and translates it to a separate, self-contained Rust crate. All crates are organized together in a Rust workspace under the folder given by the TRANSLATION_DIR environment variable, alongside the original C test_case folder:

📂C-project-name # A single C project folder with an arbitrary name
 ┣ 📂test_case
 ┣ 📂test_vectors
 ┗ 📂${TRANSLATION_DIR}
   ┣ 📂target-1
   ┃ ┣📂src
   ┃ ┃ ┣📄lib.rs / main.rs # For libraries / executables
   ┃ ┃ ┣📄wrapper.rs # Wrapper module (libraries only)
   ┃ ┃ ┗📂wrapper # Individual wrappers for every exported C symbol (libraries only)
   ┃ ┗📄Cargo.toml # For crate (translated C target)
   ┣ 📂target-2
   ┗ 📄Cargo.toml # For workspace

Inside each translated crate, there are at most two distinct components of the Rust translation:

1. A consolidated lib.rs or main.rs file containing the guaranteed memory-safe, attempted Rust translation.
2. (only for library targets/crates) A wrapper directory containing C FFI wrappers for exact backwards compatibility with the original C library.

The translation tool is also capable of testing the individual Rust crates in a workspace, but the DARPA TRACTOR evaluation schema must be followed by the .json files. See here for more details and the exact specification.

A docker image with the user-specific name ideas-${UID} can be built and directly entered using:

make docker

This allows for isolated execution in a reproducible environment.

Our translation framework treats OpenRouter as the default provider, allowing easy switching between models. The MODEL environment variable controls which LLM will be used, and should be the model's name on OpenRouter.

To run LLM-based memory-safe translation of a single project and save the translated Rust workspace in a newly created TRANSLATION_DIR sub-folder run:

make examples/C-project-name/translate \
  TRANSLATION_DIR="translated_rust" \
  OPENROUTER_API_KEY="your-key" \
  MODEL="openai/gpt-5.1" \
  TRANSLATION_TEST=smoke

If a project (library or executable) was not already found under TRANSLATION_DIR, our dependency chain will first trigger its memory-safe translation, followed by C FFI wrappers (only for libraries).

IDEAS can be used with any Anthropic model by setting the PROVIDER, MODEL, and ANTHROPIC_API_KEY variables:

make examples/C-project-name/translate \
  TRANSLATION_DIR="translated_rust" \
  ANTHROPIC_API_KEY="your-key" \
  PROVIDER="anthropic" \
  MODEL="claude-sonnet-4.6" \
  TRANSLATION_TEST=smoke

Note the prefix anthropic is missing from MODEL, and must instead be set as the PROVIDER.

IDEAS can be used with any OpenAI model by setting the PROVIDER, MODEL, and OPEN_API_KEY variables:

make examples/C-project-name/translate \
  TRANSLATION_DIR="translated_rust" \
  OPEN_API_KEY="your-key" \
  PROVIDER="openai" \
  MODEL="gpt-5.5" \
  TRANSLATION_TEST=smoke

Note the prefix anthropic is missing from MODEL, and must instead be set as the PROVIDER.

IDEAS relies on litellm, which supports many other model providers (e.g., Google Vertex, MS Azure, etc).

The instructions at https://docs.litellm.ai/docs/providers inform which parameters should be set in litellm and IDEAS flows through the dspy.LM instance.

We are working on writing a comprehensive guide for more providers, but developers can inspect how the dspy.LM is instantiated by IDEAS and infer any additional required parameters that need to be passed to litellm.

To run all tests (if available and following the DARPA TRACTOR evaluation schema) on an existing translation, run:

make examples/C-project-name/test

If a project was not already translated, this will trigger complete translation and testing, and can be used a single-click translation-and-evaluation command.

To translate, test, and aggregate statistics about all projects contained in the examples folder, run:

make -j128 examples/test \
  VERBOSE=1
  ...

This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00112590134.