Claude’s C Compiler is a C compiler written in Rust by a team of parallel Claude agents as an experiment in long-running autonomous development, described in Building a C compiler with a team of parallel Claudes. It was specified as a from-scratch optimizing compiler with no dependencies, compatible with GCC, able to compile the Linux kernel and designed to support multiple backends; the human author specified some design points, such as an SSA intermediate representation to enable multiple optimization passes, but not how to implement them.
It was built over nearly 2,000 Claude Code sessions across two weeks by Claude Opus 4.6, at a total cost of just under $20,000. The implementation is described as clean-room — Claude had no internet access at any point during development — and depends only on the Rust standard library. Its source code is publicly available.
Capabilities
The compiler is about 100,000 lines long and can build a bootable Linux 6.9 on x86, ARM and RISC-V. It can also compile QEMU, FFmpeg, SQLite, PostgreSQL and Redis, has a 99% pass rate on most compiler test suites including the GCC torture test suite, and can compile and run Doom.
Its announcement lists several limitations:
- It lacks the 16-bit x86 compiler needed to boot Linux out of real mode and calls out to GCC for that phase; its x86_32 and x86_64 compilers are its own, and on ARM and RISC-V it compiles entirely by itself. The compiler can output correct 16-bit x86, but the result is over 60kb, far beyond the 32k code limit Linux enforces.
- It has no assembler or linker of its own; those were the last parts Claude began automating and are described as still somewhat buggy, and the demo video used GCC’s assembler and linker.
- It builds many projects but not all, and is not yet a drop-in replacement for a real compiler.
- Even with all optimizations enabled, its generated code is less efficient than GCC’s with all optimizations disabled.
- The Rust code quality is described as reasonable but nowhere near what an expert Rust programmer might produce.
Adoption & Ecosystem
The compiler was built as a capability benchmark rather than as a production tool, and it is presented as having nearly reached the limits of the model that wrote it: further fixes were not fully successful, and new features and bug fixes frequently broke existing functionality. Its development relied on a harness of parallel agents coordinating through git and lock files, and on using GCC as a known-good oracle so that different agents could work on different failing kernel files at once. Readers are invited to download it and try it on their own C projects.