A harness for driving small local models through real work — writing and maintaining code, running ML pipelines, operating a service stack, and researching questions on the open web.

Install

Linux and macOS:

curl -fsSL https://codezaiku.org/install | sh

Windows, from PowerShell:

irm https://codezaiku.org/install.ps1 | iex

Both fetch the release from GitHub and verify it against the release’s own SHA256SUMS before installing anything — a mismatch refuses rather than proceeds. Only the script comes from this domain; the artifact and the checksums come from the same GitHub release, so this script cannot hand you a payload those checksums do not match. Piping a script into a shell is worth being wary of in general: fetch it, read it, then run it if you would rather.

You need a JDK 21 or newer, and an OpenAI-compatible model server — CodeZaiku bundles neither a JVM nor any model weights. On Windows you also want Git for Windows: the harness shells out through its bash. Then run codezaiku doctor, which names anything missing and the command that fixes it.

What it is

One growing conversation with good tools, built to answer a single question honestly: what does the harness contribute, and what does the model contribute? Three independent controlled experiments in the repository answer it the same way — the harness is sound; the model is the wall. Swap a 9B for a 30B and change nothing else, and the failures move.

It operates a service stack, reviews code, researches questions against live sources, and can be driven by another agent over MCP, ACP or as a plain subprocess.

Read before you rely on it

• Limitations and measured results — every number, including the ones that invalidated our own claims. Coding is capability-bound on small models; there is no long-running soak evidence yet.
• Architecture — how it works, and what was removed after measurement.
• Platforms — Linux, macOS, Windows and WSL2, with what does and does not work on each.
• Running it under another agent — the contract, the result document, and the failure modes.

Found a vulnerability? security@codezaiku.org or a private advisory — please not a public issue.