CodeZaiku
A coding and development harness that runs on local models or hosted APIs. It is an AI coding agent for your terminal, like Claude Code or aider, that works with a small model on Ollama, llama.cpp, LM Studio or any OpenAI-compatible server. It also reviews git diffs, repairs a broken docker compose stack with rollback when you allow it, reports security problems in a server's Docker containers, researches questions with cited sources, and runs as an MCP server, an ACP agent or a command that prints JSON for other agents.
Most agents like this assume a frontier model behind someone else's API. CodeZaiku is built the other way round — a 9B on hardware you own — and the harness does the work that makes that viable: deterministic evidence instead of model guesses, a guard stack around anything that acts, and verification with rollback on every change.
It runs against any OpenAI-compatible server and ships no weights. The reference tier is a 9B.
→ ARCHITECTURE.md — how it works → LIMITATIONS.md — the measured numbers
Quick start
You need an OpenAI-compatible model server, and a JDK 21+ unless you take a build that carries its own Java runtime (from 0.3.2, one per platform; the installer picks one when Java is missing). CodeZaiku ships no weights.
Linux is the supported platform. macOS runs the coding and research surfaces (and can operate remote Linux hosts over ssh); Windows works under WSL2. There is a container image. See PLATFORMS.md.
Three ways to use it
At a terminal, for yourself — carry on below; the commands are in What you can do with it.
Underneath another agent or orchestrator — CodeZaiku is a backend as well as a tool. It speaks a
CLI-subprocess contract (codezaiku run), ACP (codezaiku acp) and MCP (codezaiku mcp).
The contract, the result shape and the failure modes are in
DEPLOYING_AS_A_BACKEND.md; start there rather than here.
Underneath Wyrdsekai — CodeZaiku is released under the Wyrdsekai umbrella and is one of the coding backends Wyrdsekai can summon: you ask the companion for something, it hands the work to CodeZaiku. WITH_WYRDSEKAI.md is the short version of the page above — just the wiring, and the two things that trip people up. Nothing here depends on it; CodeZaiku runs standalone.
Installing
There is one binary artifact and it is platform-independent. CodeZaiku is JVM bytecode with no
native parts, so the same tarball installs on Linux, macOS and Windows — there is no .pkg or .msi
to look for, and none is needed.
Linux is the reference platform. macOS and native Windows are measured too, and the installer is
verified on all three from a clean unpack. On Windows, install from Git Bash — the harness shells
out through bash, and Git for Windows supplies it. See PLATFORMS.md for
per-platform steps and for what does and does not work on each.
Linux and macOS:
curl -fsSL https://codezaiku.org/install | sh
Windows, from PowerShell (you also want Git for Windows — the harness shells out through its bash):
irm https://codezaiku.org/install.ps1 | iex
Only the script comes from codezaiku.org; the artifact and the checksums both come from the same
GitHub release. What that URL serves is scripts/install-remote.sh
(and scripts/install.ps1) from this repository, verbatim — read them here
first if you would rather, and diff them against what the site serves. Both installers fetch the release tarball and verify it against the release's own
SHA256SUMS before installing anything — a mismatch refuses rather than proceeds. Only the script comes from
the URL above; the artifact and the checksums both come from the same GitHub release, so the 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.
They install to ~/.local (or %LOCALAPPDATA%\Programs on Windows). CODEZAIKU_PREFIX puts it
somewhere else, CODEZAIKU_VERSION pins a release.
No JRE on the machine yet? On Debian or Ubuntu, take the .deb from the
latest release instead — it declares a JRE
dependency, so apt installs one for you:
sudo apt install ./codezaiku_0.1.1_all.deb
The one-liners deliberately do not install a JRE themselves: a script piped into a shell should not
be reaching for sudo. They check for one; when it is missing they install the build for the platform
that carries its own runtime (codezaiku-<version>-<platform>.tar.gz, from 0.3.2; CODEZAIKU_RUNTIME=1
asks for it outright), and codezaiku update keeps such an install on its own kind.
To upgrade, run codezaiku update now, or the same install command again. update now takes the
latest release, checks it against the release's SHA256SUMS and swaps it in; the next start runs it.
When ResearchZosho is installed, it then asks ResearchZosho's own updater to do the same, and says what
came of each. Each program's own updater replaces its files, so two programs updating at once never
collide: the second finds the first at work and leaves it alone. On Windows, CodeZaiku cannot replace
its own files while it runs, so there you run the install command again. The installers resolve the
latest release each time and replace the old install rather than writing over it, so nothing stale is
left behind. Your config and data in ~/.codezaiku are untouched. The .deb upgrades in place with
sudo apt install ./codezaiku_<new>_all.deb and keeps /var/lib/codezaiku; if you enabled
codezaiku.service, the upgrade does not restart it, it tells you to when it suits you.
tar xzf codezaiku-0.1.1.tar.gz # unpacks a ready-to-run ./codezaiku
./codezaiku/bin/codezaiku --version # works immediately
sudo mv codezaiku /opt/codezaiku
sudo ln -s /opt/codezaiku/bin/codezaiku /usr/local/bin/codezaiku
On Debian or Ubuntu, sudo apt install ./codezaiku_0.1.1_all.deb — it installs to /opt/codezaiku,
links /usr/bin/codezaiku, and lets apt pull a JRE. The systemd unit it ships is disabled; nothing
starts on its own.
From source, if you want to build it yourself or work on it:
git clone https://github.com/Wyrdsekai/codezaiku.git && cd codezaiku
# 1. install — builds and puts `codezaiku` on your PATH (~/.local by default;
# --system for /usr/local, --prefix DIR for anywhere else)
scripts/install.sh
# 2. point it at your model server (install.sh already wrote a starter config)
codezaiku setup # the first ten minutes: the model (found, served on demand, or a hosted key), web search, your editor's agent over MCP, the library if wanted
codezaiku model serve install # this machine's memory, on demand: the measured model, llama.cpp behind a proxy that starts it when asked and stops it after 20 idle minutes (on Linux llama.cpp's container for the GPU it finds: NVIDIA, AMD or Intel; Metal on macOS, Vulkan on Windows)
codezaiku model detect # finds Ollama / llama.cpp / LM Studio / vLLM if running
codezaiku model use http://localhost:8200
codezaiku model add deepseek openrouter deepseek/deepseek-v4.1-flash --prices 0.032/0.144/0.002 # a hosted model under a name,
codezaiku model key openrouter --file ~/openrouter-key.txt # its key kept for its host,
# chosen per job: codezaiku code . "…" --model deepseek; `codezaiku spend` says what the jobs cost
# 3. check the environment — names exactly what is missing and how to fix it
codezaiku doctor
# 4. confirm the model answers and can call a tool
codezaiku smoke
The install carries its own launcher, jars and knowledge library, and downloads nothing at runtime.
It does not bundle a JVM — install.sh checks for a JDK 21+ and stops if it does not find one.
No model weights are bundled either. scripts/install.sh --uninstall removes the program and
keeps ~/.codezaiku — your config, learned cards, research findings and audit trail.
--purge removes those too, after listing what will be lost and asking you to confirm.
To work in the repo without installing, bin/codezaiku runs from the checkout.
If you have no server yet:
docker run -d --name codezaiku-drive -p 8200:8200 \
-v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:server-cuda \
-m /models/<your-model>.gguf --port 8200 --host 0.0.0.0 --jinja --ctx-size 32768
To use the models in your own AWS account (Amazon Bedrock) with your own AWS sign-in, see BEDROCK.md.
--jinja is required — without it the model returns tool calls as prose and nothing works.
See MODELS.md for what the harness needs from a model, what we measured on,
and why codezaiku smoke is the check that matters.
Debian / Ubuntu
sudo apt install ./codezaiku_0.1.1_all.deb # build it: packaging/deb/build-deb.sh
Installs to /opt/codezaiku with /usr/bin/codezaiku. A systemd unit is included but not
enabled — CodeZaiku can modify live systems, so starting it is a deliberate act.
Verifying what you downloaded
Every release asset ships with SHA256SUMS and a Sigstore bundle (<asset>.sigstore.json):
sha256sum -c SHA256SUMS --ignore-missing
gh attestation verify codezaiku-0.1.1.tar.gz --repo Wyrdsekai/codezaiku \
--predicate-type https://codezaiku.org/attestation/release/v1
gh attestation needs GitHub CLI 2.49 or newer. An older gh reports unknown command with no hint why — check with gh --version before concluding the signature is bad.
The signature says the release workflow, running at that tag, blessed those exact bytes. It is an authenticity statement, not build provenance — the artifacts are built and validated on real hardware rather than in CI, because an artifact nobody ran is not one worth shipping. Release assets are immutable; a fix ships as a new version, never as a re-upload.
What you can do with it
codezaiku help # all commands, grouped by intent
Talk to it
codezaiku chat ~/myproject
A coding assistant you talk to in your terminal — the same kind of tool as Claude Code, Codex CLI or aider, except the model behind it is yours to choose: one running on your own hardware, or a hosted API. You type what you want — "why does the login test fail?", "add a retry to the uploader", "find out how other projects handle this" — and it reads your code, makes changes, runs commands, or searches the web, then reports back. You steer it turn by turn.
By default it asks before changing or running anything, and shows you the exact change first.
You answer once, or "always" for that command. /undo reverses what it did, step by step.
Conversations are saved and can be resumed later, and /remember keeps a fact available in
every future conversation in that project. /switch <project> moves the chat to another project in the same terminal and tells it where you
were. /help inside the chat lists everything else —
background tasks, delegating work to a second agent, switching models, cost tracking.
codezaiku v1 ~/myproject serves the same assistant as an OpenAI-compatible API
(/v1/chat/completions), so a chat UI like Open WebUI can be the front end. Over that wire it
can only read, never change anything — an API cannot ask you for permission, so it is not given
anything that would need it.
codezaiku web is the chat in the browser, for several projects under one server: open
http://127.0.0.1:7072/, pick a project or add one by its folder, and talk. It is the same chat as
the terminal's, so it asks before it writes or runs anything, and you answer with buttons on the
page. Each project keeps its own conversation. Moving to another project is a click, and "switch"
also carries where you were, so the next thing you say there starts from it. It listens on this
machine only unless you say --host 0.0.0.0, which opens it to your network; it then prints a token
and the addresses to type, and the browser asks for the token once. From the menu's People page you
can name others: a viewer sees every page and every chat as it runs and can type nothing, an owner can
do all you can. Each signs in with a name and a password.
Research a question
codezaiku research "your question" broad
codezaiku research "your question" fan
It searches the web, reads pages, and writes an answer that cites its sources. broad is one
researcher working alone. fan splits the question into parts, researches them in parallel, has
a critic check whether the coverage is actually sufficient, and then writes the combined answer
— slower, and much better on questions with many parts.
Search works out of the box, and works better with a backend you configure. Without one, web_search
falls back to Wikipedia plus the scholarly literature (Crossref and OpenAlex): reference pages and papers by
DOI, not the whole web. Every research run also has scholar_search, the literature by DOI, whatever the
backend. For the whole web, two options, the first recommended:
codezaiku config set CODEZAIKU_BRAVE_KEY <key> # recommended: Brave Search API
codezaiku config set CODEZAIKU_SEARXNG http://host:8888 # or: your own SearXNG instance
The Brave Search API has a free tier (2,000 queries/month) and
gives markedly better results — a search tool is only as good as what it searches. SearXNG is
fully self-hosted and needs no account, but it aggregates public engines that rate-limit under
sustained use. With both configured, Brave is used first and SearXNG is the fallback. Setting
either one is also what enables web search inside chat.
What a research run finds is kept. The research memory in ~/.codezaiku/research holds each
finding with its source, and the next run on a related question starts from it instead of from
zero. That is the built-in tier: it works with nothing else installed.
For real research, the kind that takes hours and comes back with a report you can check, there is
ResearchZosho, a separate program from the same family. codezaiku install researchzosho fetches it, verifies it and runs its setup. From then on the chat's
/research files runs with it, /librarian asks what it holds, and what it holds is pushed into
every turn. The questions in the research memory are handed to it as open questions. codezaiku doctor
and the chat say when a newer ResearchZosho is released; codezaiku update now has ResearchZosho's own
updater install it, and the library and settings stay.
Work on a codebase
codezaiku code ~/myproject "add pagination to the /items endpoint"
codezaiku decompose ~/myproject "add an admin dashboard with login and audit view"
codezaiku review ~/myproject HEAD~3 # read-only review -> findings
code keeps one conversation going until the goal is met. Use it when the model can hold the
whole job at once. decompose splits the goal into ordered slices first, then runs the same
loop with that plan pinned as a TODO it re-injects every turn — so you have a running program
after the first slice, and it grows from there. Use it when the job is big enough that the
model would otherwise lose the thread. loop is an older name for code and still works.
Operate a stack — sense, localize the failing service, fix, verify, roll back on failure
codezaiku fix myproject propose # diagnose and propose (default, safe)
codezaiku fix myproject guarded # act, behind the full guard stack
codezaiku watch myproject propose 60 # continuous: fix on degraded, else report
codezaiku investigate myproject # diagnose only — capped at localize, never acts
Scope can be local (myproject), remote (ssh://host/myproject) or a container
(docker://name). Everything else — the health endpoint, the verify command — is auto-discovered.
Operate a whole machine — the same operator, with the boundary set to the box instead of one stack
codezaiku triage local # enumerate the box, report what is broken
codezaiku triage local "" 20 guarded # ...and repair the unhealthy compose stacks
codezaiku watch-machine local 300 # the same sweep on a loop, as a daemon
triage enumerates everything it can see — compose stacks, failed systemd units, host disk
and memory — runs the SRE operator against each unhealthy stack, and surfaces the systemd and
host findings for a human. Each stack is still bound by blast-radius and the authority ladder,
so widening the boundary does not widen what any single fix may touch.
Note the difference in defaults, because it is the one place where a whole-machine command is
less conservative than a per-stack one: triage reports unless you pass guarded or
unattended, while watch-machine defaults to guarded — it is a daemon meant to keep a
box healthy, so it acts. Pass a ceiling explicitly if you want it to only watch:
codezaiku watch-machine local 300 "" propose. Run it under systemd on a real box.
Review security posture — report only, never auto-remediates
codezaiku secure local
Research a question against live sources
codezaiku research "what changed in the Foo API in v3" depth
Drive it from another agent
codezaiku run --text "add pagination" --output-format json # one task, one JSON result
codezaiku acp # Agent Client Protocol v1 agent on stdio
codezaiku mcp # MCP server on stdio, all surfaces exposed
npx -y @wyrdsekai/codezaiku-mcp # the same, from any client that runs npm packages; installs the release if needed
codezaiku serve 7070 # HTTP: POST /fix, GET /health, GET /audit
Safety
The operator can change your systems, so the defaults are conservative and the guards are not optional.
- Authority ladder —
observe<localize<propose<guarded<unattended. Defaultpropose: it tells you the command, it does not run it. An unproven fix never auto-applies. - Blast radius 1 — a fix may touch the service it localized and nothing else.
- Rollback — snapshot before a destructive change, verify after, restore if verification fails. This has a counterfactual: with it on, a model that corrupted a database config was rolled back and the stack came back green; with it off, the same failure left the service dead.
- Harm check — a fix that repairs its target but breaks a healthy dependency is a failure.
- Kill switch —
CODEZAIKU_OPS_HALT=onforcesobserveimmediately. - Security findings are report-only at every rung. A wrong reliability fix restarts a service; a wrong containment action firewalls your own load balancer.
Untrusted input is treated as untrusted: container logs are attacker-writable, and a measured injection through them steered the operator 40/40 on both model tiers before it was fixed structurally. See ARCHITECTURE.md §6.
What this is not
- Not a frontier coding agent. SWE-bench Lite ≈1/8 on the 9B; refactoring 0/8. It is not a replacement for tools built on much larger models.
- Not production-hardened. No long-running soak evidence yet.
- Not a multi-agent framework. One loop, good tools, hard measurement. An earlier blackboard/agent architecture was designed, partly built, and removed.
The full accounting is in LIMITATIONS.md.
Repository layout
bench/ is shipped on purpose. The rigs are how every number in LIMITATIONS.md was produced, including
the ones that came out badly.
Documentation
- ARCHITECTURE.md — design
- LIMITATIONS.md — the measured numbers
- MODELS.md — choosing and running a model server
- PLATFORMS.md — Linux, Docker, macOS, Windows
- CONFIGURATION.md — the six settings that matter, and the rest
- DEPLOYING_AS_A_BACKEND.md — running it under another agent or orchestrator
- WITH_WYRDSEKAI.md — wiring it up as a Wyrdsekai coding backend
- ROADMAP.md — what is next, ordered by what measurement says is holding it back
- CONTRIBUTING.md — including the measurement rules
- SECURITY.md — reporting, and the agent's own threat model
Getting in touch. Bugs and feature requests belong in issues, where other people can see them and the answer. For anything else, [email protected]. Vulnerabilities go to [email protected] or a private advisory — see SECURITY.md; conduct concerns to [email protected].
The family
CodeZaiku is part of Wyrdsekai, and its sibling is
ResearchZosho, the research library: it keeps what research runs find,
checks the claims against their sources, and answers questions from what it holds.
codezaiku install researchzosho puts it beside CodeZaiku, and the two share a model server, so one
on-demand model on a machine serves both.
License
Apache 2.0 — see LICENSE.
