The AI agent ecosystem is drowning in bloat. While most developer-facing agents rely on heavy runtimes like Node.js or Python, a new project called agentc is going the opposite direction. Built in freestanding C23 with zero libc dependencies, agentc compiles to a static binary under 1MB that starts in less than a millisecond. It’s a direct challenge to the notion that you need a full runtime environment to interact with LLMs.

Bare-Metal Efficiency

The developer behind agentc, inspired by the pi coding agent, aimed for extreme minimalism. The release binary comes in at approximately 760 KiB stripped, with an idle RSS footprint of just 0.7 MB. This makes it viable for embedding in other binaries or running on constrained hardware where traditional agents would fail. The toolchain uses clang with -ffreestanding and -nostdlib, implementing its own allocator and entry point to avoid any host libc leakage.

Cross-Platform Without the CRT

Despite its minimal footprint, agentc supports a wide range of targets including Linux x86-64, aarch64, and riscv64, as well as macOS and Windows. The Windows build links no CRT or SDK, using import libraries generated from definition files. For TLS, the project vendors a freestanding version of mbedTLS 3.6.2 on Linux, while relying on OS-native stacks like SecureTransport and SChannel on Apple and Microsoft platforms. This approach ensures the binary remains self-contained and portable without dragging in unnecessary system libraries.

Extension Model and MCP Support

agentc isn't just a toy; it supports the Model Context Protocol (MCP) for tools, prompts, and resources. It features a versioned C ABI for extensions, allowing developers to write plugins in C or Rust. These extensions can be linked statically or loaded dynamically on macOS and Windows. The agent includes a robust provider system with native support for Anthropic, OpenAI, Google, and local Ollama instances, alongside presets for OpenRouter, xAI, and others. Model discovery is cached locally to minimize network overhead.

Key Takeaways

  • Sub-1MB static binary with <1ms startup time and 0.7MB idle RSS.
  • Freestanding C23 implementation with no libc, using vendored mbedTLS on Linux.
  • Supports MCP tools, prompts, and resources with a versioned C ABI for extensions.
  • Cross-compiles for Linux, macOS, and Windows without relying on standard CRTs.

The Bottom Line

While the HN score is low, this project represents a necessary correction in AI tooling. If we want agents to run everywhere, they need to stop assuming every machine has a gigabyte of RAM to spare for a runtime. agentc proves that high-performance, embeddable AI interfaces are possible in pure C.