FlosForge Toolbox is on the Mac App Store

FlosForge Toolbox is out, and it is the first thing I have shipped on the Mac App Store. Everything else on this site so far has been a GitHub download or a set of build steps. This one you can install the way you install any other Mac app.

It is a free download with a 7-day trial that unlocks everything. When the trial ends nothing is charged: the app asks you to pick Lifetime, a one-time purchase, or Yearly, a subscription. There is no account, and purchases stay between you and Apple.

What it is

Toolbox is the drawer of small utilities every developer reaches for several times a day: Base64 and URL encoding, JWT decoding, hashes and HMACs, JSON and SQL formatting, regex testing, cron parsing, UUIDs, timestamps, DNS lookups, port checks. That is 420 tools in 17 categories, in one native window. Four hundred of them never touch the network.

The reason it exists is simple. Most of us do these jobs by pasting things into whichever website comes up first, and the things we paste are tokens, payloads, log lines and customer records. Toolbox does the same jobs on your Mac, and nothing you do is reported anywhere.

Where it gets interesting: prompts as tools

The part I use most is not one of the 420. It is the AI Prompts section, which turns a prompt you would otherwise type into a chat window again and again into a tool that sits in your library next to the others.

You write the prompt once, with placeholders in double braces:

Rewrite this release note for {{audience}}, in at most {{sentences}} sentences:

{{input}}

Each placeholder becomes a field in the tool’s form. You decide what kind of field: a line, a paragraph, a number, a yes-or-no, or a pick-one list of values you name, so {{audience}} can offer “customers”, “developers” and “my manager” and nothing else. Save it, and it is a tool: it has presets, it keeps a run history, and it can be found from the ⌘K palette like everything else.

{{input}} is the special one. It never becomes a field. It is filled in automatically when the prompt runs as an action on another tool’s result. That is what lets these prompts slot into the work you are already doing:

A DNS lookup comes back with something odd. Write a short prompt that explains DNS answers, attach it to the DNS tools, and it is one click away on every result, without leaving the tool.
A regex you inherited. Paste it into the regex tester, then run Explain Regex on it, or ask Write a Regex for the one you actually need.
A stack trace from a log. Explain This Error, then Rubber Duck if the answer raises more questions.
A diff you are about to commit. Commit Message drafts the message. You still decide whether it is right.
A JSON response. JSON → Type turns it into a type definition, and curl → Code turns the curl command that fetched it into code.

Fifteen of these come installed, and they are ordinary prompts once they are there: edit them, rename them, or delete the ones you never use. The ones you write yourself are first-class too. You can make one appear as its own tool, as an action on other tools’ results, or both.

The model runs where you choose

Every prompt tool runs on one of three providers, and you tick which ones each tool may use:

Apple Intelligence, on-device, on a Mac that supports it. No key, no setup.
Ollama, on this Mac, or on another Mac on your network that has a bigger GPU.
Google Gemini, only once you paste in your own API key. It is the only provider that sends anything off your Mac, and it goes straight to Google, never through a server of mine. There isn’t one.

With the first two, a prompt about your production logs never leaves the building. That is the whole point of putting AI in a local-first app. A cloud chat window would be easier to build, and it would undo everything the rest of the app is careful about.

The app is also careful about the runs you did not start yourself. A tool that recomputes as you type never reaches a hosted provider at all. A result sent on from another tool asks before it goes to Gemini, or stays on-device if you prefer. Every AI call is recorded in a ledger: the time, the provider, the model and the token counts, never the prompt itself. You can set token limits per hour, day, week or month, and on-device models never count against them.

Getting it

Toolbox needs macOS 26 or later. The Toolbox page has screenshots and the full list of what is in the box, its privacy statement lists every connection the app can make, and the changelog covers 1.0.1 and 1.0.2.

If you have been pasting secrets into random websites, try it for a week and see whether you go back.