Every AI product on the market claims to be secure, private, and safe. Those are adjectives. They ask for your trust and offer you their word.
I built my project on a different bet: that trust should not be a promise at all. It should be a mechanism you can watch working.
This week I published the first public piece of it — a free, open-source command-line toolkit — and I'd like to show you what's in it, how it's built, and what it's a sampler for.
The free toolkit
Wizard Central Toolkit (Apache 2.0) is a local-first CLI suite: a router plus 12 "wizards" — small, single-purpose tools that each do one job well.
pip install wizard-central-toolkit One naming rule runs through the whole project, and it matters more than it looks:
A command is what you type. The name is what the project is.
So each wizard has a full canonical name — who it is — and a command — what you type. Both spellings of the command run: the short form (codeguardwiz) and the full name-command (codeguard_wizard).
| Canonical Name | You type | What it does |
|---|---|---|
| Wizarding Router | wiz | Routes what you type to the right wizard |
| Godfather Wizard | godfatherwiz | Fleet dispatcher + interactive REPL (39 slash command forms) |
| Code Guard Wizard | codeguardwiz | Static code audit, plain verdicts, optional local-model seat |
| Python Wizard | pythonwiz | Parse-only Python checks; secrets scan that redacts by default |
| Write Wizard | writewiz | Consent wall for file writes: propose → preview → commit, rollback |
| Read Wizard | readwiz | ALLOW / DENIED policy verdicts on raw shell commands |
| Debug Wizard | debugwiz | Traceback autopsy, error census, CI-ready JSON reports |
| Diff Wizard | diffwiz | File/directory diffs, word-level diffs, honest verdicts |
| JSON Wizard | jsonwiz | Validate, pretty, compact, inspect, merge, structural diff |
| Banner Wizard | bannerwiz | ASCII banners — 571+ fonts found at runtime, 18 styles |
| Todo Wizard | todowiz | A personal work ledger: add, list, done, overdue, archive |
| Tmux Wizard | tmuxwiz | tmux session management |
| Rust Wizard | rustwiz | Rust file/project checks and crate builds |
147 subcommands in all. The whole suite is Python standard library, apart from two small display libraries — no heavy stack to audit, install, or trust.
Three rules the toolkit already lives by
Consent before change. The Write Wizard never just writes. It proposes a change, shows you the preview, and commits only on your yes — with backups and rollback if you change your mind. The Read Wizard exists for the opposite direction: before a shell command runs, it can return a plain ALLOW or DENIED.
Redaction by default. The Python Wizard's secrets scan reports the file, the line, and the kind of secret — never the value. Printing the value is a deliberate opt-in flag, because a scan you paste into an issue tracker should not become the leak it was hunting.
Honest verdicts. The tools say what they found in plain words, including when the answer is "this doesn't exist" or "I can't do that here." I care about this more than features, and it leads directly to the honest note below.
An honest note about hardware
Code Guard's static audit runs anywhere. Its optional local-model seat (via Ollama) is a different story: it needs a machine that can run a small model at a reasonable speed. I learned this the hard way — a raw audit prompt on my low-end Chromebook (Celeron, 2.7 GB RAM) ran past ten minutes unfinished, while a phone completed the same audit in about a minute. Guidance fixes confusion; it does not fix arithmetic. On a small machine, the static engine is the supported path, and the toolkit says so.
Caveats, up front
This is a young project — v0.2.8, one developer, first public release. The test suite stands at 561 passing tests, and the package on PyPI is exactly what the repo builds, but you may still find rough edges. If you do, I want to hear about them. Issues, criticism, and "why did you do it this way?" are all welcome — that's how it gets better.
What's coming next
The free toolkit is the sampler for the bigger system: Wizard Central Pro AI Modes Toolkit — a gate between a person and their AI models.
The design in one paragraph: every tool call an AI makes crosses one checkpoint, the MCP gate. The gate is default-closed and opened only by the person. The model is offered just 75 of the system's 349 commands; the rest can never be reached by a model at all. Access is granted folder by folder and expires. Destructive actions ask twice. And every crossing — served or refused — is written to a hash-chained receipts log, so the history can be checked, not just believed.
Around that gate: 39 of 41 wizards are built; 70 AI modes advise (prompts the model may follow), while guard modes bind (server code that refuses) — 9 of 23 guard modes are built so far. Still to come: the final two wizards, seat cards (saved model personas), a creation system of nine kinds that lets users build their own wizards and modes, 17 model providers, and the interactive REPL running through the gate.
When — and only when — the build is complete, Pro opens to a closed beta of 100 testers, in waves. No dates. It opens when it's finished, not before.
Why build the gate at all? Because the model is not the authority. Wizard Central does not ask the model to be good. It asks the model to be useful, and it keeps the authority where authority belongs: the model drafts, researches, and reasons; the person grants, reviews, approves, and answers for the result.
Try it
- PyPI: https://pypi.org/project/wizard-central-toolkit/
- GitHub: https://github.com/rewardsbycode-eng/wizard-central-toolkit
pip install wizard-central-toolkit wiz --help Thanks for reading. I'll be in the comments — tell me what broke, what's missing, and what you'd build with a gate like this.
A command is what you type. The name is what the project is.
