agentc

agentc is an open-source coding agent from the abird-ai organization, written in freestanding C23 with no libc. It ships as a static binary of about 940 KiB stripped and runs in a terminal, either as a TUI or in one-shot mode. The latest release listed is v0.6.0.

What It Is

agentc is a terminal coding agent that talks to LLM providers, calls tools, and keeps sessions as JSONL transcripts. It uses its own allocator and startup code, vendored mbedTLS on Linux, and the native OS TLS stacks on macOS and Windows. The same source tree builds for Linux (x86-64, aarch64, riscv64), macOS (arm64, x86-64) and Windows x86-64.

A typical flow is to run agentc setup to pick a provider, store credentials and choose a model. Then run agentc -p "..." for a one-shot prompt, or start the TUI without -p.

Providers and models

Built-in providers include Anthropic, OpenAI, Google Gemini, Ollama (local and cloud) and OpenAI-compatible presets such as OpenRouter, xAI, DeepSeek, Groq, Mistral and Together. Model lists combine a built-in catalog with endpoint discovery, cached for 24 hours. Subscription OAuth logins are supported for Anthropic and OpenAI, and custom OpenAI-compatible gateways can be added through config.

Extensibility and MCP

agentc supports the Model Context Protocol for tools, prompts and resources. MCP prompts become slash commands, and resources are read through generic read-only tools. Beyond MCP, a versioned C ABI (include/agentc_ext.h) lets C and Rust extensions link into the binary. On macOS and Windows, extensions can also load at runtime.

Interactive use

The TUI offers slash commands for switching models and thinking level, resuming sessions, compacting context, theming and running skills or prompt templates. It has inline, scrollback and fullscreen modes. Parallel tool execution is on by default.

Footprint and platform notes

The README's own measurements on one Linux x86-64 machine report about 0.7 MB cold-start RSS and sub-millisecond launch, compared against codex, Claude Code and pi. The README states that macOS and Windows builds are verified in CI but have not been driven by hand. It also notes that Windows TLS against a real provider is unverified.