NeverC

Design goals

NeverC is presented as a C23 compiler built on LLVM with a focus on security research and on being friendly to language model generated code. The README says it compiles standard C into hosted binaries, freestanding executables, and position independent dyncode from a single toolchain, targeting x86_64 and AArch64 in little endian only, with future plans to add EVM for Ethereum smart contracts and Solana eBPF for on chain programs as compilation targets. The stated design is pure C23 with no templates, no RAII, no operator overloading, and no hidden control flow, so what is written is what runs. A built in string type offers value semantics with plus, equality, and starts_with operations and automatic cleanup, avoiding the need for C++. Error handling stays explicit with no exceptions, which removes stack unwinding and the performance surprises that come with it. The project ships as a single binary that includes the compiler, linker, and runtimes with zero external dependencies to set up. The README argues that a minimal grammar and deterministic semantics make AI generated NeverC code compile correctly more often than C++ alternatives, which is the sense in which it is AI friendly. True cross compilation is listed as a goal: from macOS or Linux you can build Windows PE, Linux ELF, macOS Mach-O, Android ELF, and dyncode, with the platform SDKs bundled inside the compiler so there is no VM, dual boot, or SDK hunting.

DynCode and built in runtimes

A distinctive part of NeverC is its DynCode compiler, described as a multi stage IR and MIR pipeline with cross platform extraction, import and syscall lowering, kernel mode support, bad byte auditing, and a plugin architecture. DynCode is the mechanism for producing position independent, injectable code, which is why the project positions itself for security research rather than only general application building. The integrated linker handles COFF, ELF, and Mach-O in one binary, removing the need for an external ld or link.exe. The built in runtimes are LLVM bitcode embedded in the compiler: a string runtime with value semantic strings and dot call methods and automatic memory management, a mimalloc runtime that transparently overrides the allocator and is on by default outside kernel and freestanding targets, an xorstr runtime that does per instance compile time string encryption with mandatory late sealing and per call native expansion, and a strhash runtime for compile time string hashing with a matching runtime. The xorstr and strhash runtimes are the pieces most tied to the security story, since they obscure string constants at compile time. The README also mentions release stripping through a built in strip flag that removes non runtime symbols and source level debug information, including kernel aware dot ko structural symbol renaming that is described as a rename rather than a hash or encryption. These features are presented as native to the toolchain rather than bolted on through external scripts.

Plugin and extension model

NeverC exposes a plugin API meant for out of tree extensions. The README describes it as a pure C ABI for plugins, with a single header SDK that has zero LLVM or CRT dependencies and spans driver, preprocessor, AST, IR, MIR, MC, object, link, LTO, and dyncode phases, so a plugin can intercept any stage from IR optimization to final binary output without LLVM knowledge. A .nc file extension is offered to auto enable all NeverC features, including the string type and Rust style integer types, without extra flags. The project also ships a lean LLVM build that keeps only the x86_64 and AArch64 backends and strips the C plus plus, Objective C, and OpenMP paths, which shrinks the surface and the build. The README notes that security research features such as DynCode compilation, compile time string encryption, and cross platform PE generation are native to the toolchain rather than afterthoughts. Because the project is marked work in progress, some of the described targets and features are forward looking, and a reader should treat the current release as early rather than complete. The breadth of the plugin hooks, from preprocessing to linking, is what would let a researcher adapt the compiler to a custom payload or analysis without forking LLVM directly. The repository is published under the AGPL-3.0 license and its most recent commit was on 2026-08-26.

Editorial conclusion

NeverC is described as the AI friendly C23 compiler for security research, built on LLVM. The README says it compiles standard C into hosted binaries, freestanding executables, and position independent dyncode from one toolchain, targeting x86_64 and AArch64. It ships an integrated linker for COFF, ELF, and Mach-O and built in LLVM bitcode runtimes including string, mimalloc, xorstr, and strhash. The repository is published under the AGPL-3.0 license and its most recent commit was on 2026-08-26.