Your AI co-translator for automating C-to-Rust work.
Why · Install · Usage · How it works · Results · Paper
Crustify is a plugin for Claude Code and Codex that turns LLMs into specialized C-to-safe-Rust translators. It automates:
- Migrations. Port whole codebases or a few components that keep working with the remaining C, and offer an API compatible with both Rust and C consumers.
- Wrappers. A safe, ergonomic Rust API over C you keep.
- Audits. Undefined behavior, unsound contracts, divergences from the C and other bug classes in Rust that meets C, each with a reproducer.
- ✅ Verified empirically in an agentic feedback loop: independent reviewers, tests that pass Miri and the sanitizers, lints and formatters. Every fix comes with a regression test.
- ⚡ Fast Rust code, never at the cost of safety or correctness.
- 🎯 Equivalent to the C, without its defects.
- 🔍 Correctness gaps found against whatever spec is available: documentation, headers, tests, reference implementations.
- 🧩 Customizable with further objectives in the prompt, such as matching the C's style or maximizing coverage-guided testing.
- 📈 Improves as it goes: recurring defect classes become checks that run on every later component and stay with your code.
- 🕹️ Autonomous or supervised: let it run end to end, or step in at any point.
Crustify runs on 64-bit Linux. It needs Claude Code or Codex, and Python 3.13. Clone the
components side by side and install the tools, each on its devel branch; ffibox is a
crate the port depends on, so it is only cloned.
git clone https://github.com/crustify-rs/crustify.git
git clone --branch devel https://github.com/crustify-rs/ffibox.git
git clone --branch devel https://github.com/crustify-rs/ffimap.git
python -m pip install -e ffimapTip
To skip the setup, templates/docker/ builds an image with
everything a port needs: both agent CLIs, CodeQL, bindgen, Miri and the sanitizer
toolchains.
Open Crustify in the checkout of the C codebase, and name the task in your first message: the components to port, the build configuration, and anything else the port must respect.
Claude Code
claude --plugin-dir /path/to/crustify --agent crustify:crustifyCodex
Install the profile once, then open it:
python3 /path/to/crustify/providers/codex/install.py # add --model-proxy openrouter --model <model> for OpenRouter
codex --profile crustifyRerun the installer after the checkout changes. It warns when Codex's sandbox cannot start
on the host; --sandbox full is for a container.
Other agent harnesses
providers/generic/compose.py prints the whole prompt,
with an index of the skills, for any harness that can run shell commands, edit files and
spawn sub-agents:
python3 /path/to/crustify/providers/generic/compose.pyOne agent runs the port. It builds the C, decides the scope and the order, spawns
subagents to port independent components in parallel and others to review them, and
merges their work. The system prompt,
agents/crustify.md, gives the requirements of the ported code,
functional and non-functional, and the process: work in parallel, review every component
against every requirement, learn from earlier findings, and keep STATUS.md current. Everything else is a skill:
| group | skills |
|---|---|
conventions/ |
the port's files under .crustify/, the anchors that name the C item each Rust item stands in for, the lints, and the language of its texts |
tools/ |
bindgen, ffibox, ffimap |
The port ends with .crustify/REPORT.md: the configuration, the boundary between Rust and
C, the defects found in the C, the performance numbers and what was verified how.
Crustify's earlier, multi-agent version generated safe bindings for two libraries, with little guidance from one developer overseeing it, as our paper details. Binder and reviewer agents mainly ran Claude Opus 5.5; GPT-6.1 Sol and GPT-6 Astra wrapped a few modules, and Claude Fable 5.1 ran the early reviews.
| public API items wrapped | Rust line coverage | unsafe public API |
UB defects fixed in review |
| libgit2 | FFmpeg | |
|---|---|---|
| Public API wrapped | 480 of 1,146 items (41.9%) | 573 of 1,338 items (42.8%) |
| Wall-clock time | 51.9 h | 58.9 h |
| Cost | $2,308 | $2,120 |
unsafe public API (consumer-facing) |
42 of 2,663 functions (1.6%) | 11 of 2,637 functions (0.4%) |
| Raw pointers in public API (consumer-facing) | 2 functions (0.1%) | 16 functions (0.6%) |
| Tests: unit / equivalence / UB | 787 / 489 / 788 | 968 / 723 / 1,003 |
| Line coverage, Rust / C | 98.9% / 46.0% | 99.1% / 67.2% (libavutil) |
| UB defects the reviewers fixed | 8 | 13 |
| Rust LoC | ~18K | ~19K |
Each repository holds the code, the campaign's plan, its review reports and its dashboard.
agents/crustify.md: the system promptskills/: the skillsrsi/: the patterns ports learn that hold beyond their codebase, and the harness changes they proposeproviders/: the prompt composed for each host:claude/,codex/, andgeneric/for any othertemplates/: the format ofSTATUS.md, and the Dockerfile
- Marius Momeu, University of California, Berkeley
@misc{momeu2026crustify,
title = {Crustify: Repository-Level Generation of Safe Rust Bindings with {LLM} Agents and Coverage-Guided Testing},
author = {Momeu, Marius and Arvinte, Marius and Cornelius, Cory and Sen, Koushik},
year = {2026},
eprint = {2611.10394},
archivePrefix = {arXiv},
primaryClass = {cs.PL},
url = {https://arxiv.org/abs/2611.10394}
}