Today marks the public release of jichi. A project that grew out of an experiment, and a few questions: Where is all this software people are developing with AI assistance? How do I learn developing software with or without AI support? And how long is the journey?
From the documentation a few snippets:
A complete AI coding agent in C, conforming to the C89 (ANSI C / C90) standard, for Linux/POSIX: chat with any LLM, an agentic tool-use loop, retrieval, snapshots and undo, a bounded autonomy envelope, MCP/LSP/ACP integration, an interactive terminal UI and a scriptable headless mode — one small binary that links libcurl and nothing else.
jichi is Japanese. Written 自治（じち） it means autonomy / self-government — 自 (self) + 治 (govern) — an everyday word, as in 自治体（じちたい） (a municipality). That is the sense meant here: this is an agent you hand a bounded goal to, with a token and wall-clock budget, an edit-scope fence and a verification gate, and it governs itself inside those limits. A rarer, literary homophone 自知（じち） reads as self-knowledge, which suits a tool that keeps its own run journals and mines them to correct itself.
jichi compiles under four toolchains — gcc and clang (the CI-gated
pair, strict -std=c89 -pedantic), g++ (as C++17; make CC=g++, see
docs/CPP_BUILD.md), and zig cc (make CC="zig cc", including a fully static -target x86_64-linux-musl build — see
docs/ZIG_BUILD.md). The shipped build remains
gcc/clang C89; the other two are maintained properties of the codebase.
ARM cross-builds ride the same zig cc -target path — one command each,
zero warnings, for the single-board Linux targets (Raspberry Pi, Arduino UNO
Q; see docs/plans/2026-07-hardware-testing.md).
Why you compile it yourself. jichi is distributed as source code, not as a
ready-made program. You run a single command (make) that turns the source into
the two programs, on your machine, for your machine. This keeps the whole thing
small, auditable, and dependency-light — and, if you are learning, building it
is the first exercise of the curriculum: a real program, a real toolchain, a
real green result at the end.
