FB/010.5.1
The workbench
FeatDesk
Markdown notes in a folder you own, in three states: capture, bench and archive. Search everything; the app never touches the internet.
Tool suite0.5.1 · pilot open
Work with AI without losing control.
Featbox is a suite of tools for people running long projects with several AI assistants. Every AI has its role, and you have the final word.
[01] The approach
Method before tools.
AI without a process is automated chaos. Featbox starts from three principles, and the tools exist to keep them.
01
One role for each AI
Each assistant does what it does best: planning, research, review. The project context is the same for all of them.
02
Humans decide
The AI proposes, researches and executes. What becomes a rule goes through a person's approval and is recorded with date and source.
03
No provider sees everything
Each provider gets only what its role needs. The full project stays with you, in your own files.
[02] The tools
Three tools, one way of working.
The names follow a simple rule: Feat + the job. The screens below are from the installed 0.5.1, with demo data.
FB/020.5.1
The reader
FeatReader
Reads md, txt, pdf and docx. Each document gets a notebook with highlights, notes and audio, and text-to-speech runs on your computer. Free.
FB/030.5.1
The decision room
FeatRoom
Every request, approval and rejection goes into a hash chain: you can prove what was decided and when. One personal chain and one per project.
[03] How a decision moves
The AI proposes. You decide.
Alongside, an illustration of how a proposal becomes a rule.
- proposalan AI suggests a change, with its source
- reviewanother AI, with another role, reviews the proposal
- waitingyou approve or reject
- approvedit becomes a rule, with date and source recorded
[04] How it is built
The same standard the product stands for.
Featbox is built with AI, with a person approving what matters.
01
Tested and documented
Nothing is called done without automated tests and a record of why each decision was made.
02
Decisions in writing
Architecture, decisions and tests are written down in every project.
03
Open showcase, private code
This page and the presentation docs are public. The code stays private.
[05] In practice
What we can already show.
Only what works and can be checked, or what was measured, goes here. Anything still a hypothesis stays out.
Nothing leaves your computer.
No feature uses cloud AI in this version, and each app's About screen says so. FeatDesk blocks every network connection. FeatRoom's language model is optional, off by default, and only accepts a model running on your own PC.
Only what you approve becomes a rule.
The AI logs a decision request; you approve or reject it. Each event type has its own color, and a rejection is recorded just like an approval.
A history that can't be erased.
Each event stores the SHA-256 of the previous one, in a file that only grows. The app checks the chain and shows "chain intact". If anyone edits an old record, the check flags it.
The why stays next to the code.
Each module of the project has a short card describing what it does. When new code lands after it, the badge changes from "up to date" to "outdated", so you know what the AI should no longer trust.
What working with AI really costs.
One real week of work (Oct 17–23, 2026): about 650 million tokens across 33 sessions. On average, the AI re-read 380k tokens on every reply. The biggest cost comes from long conversations, not from code.
cost meter in testing; joining the suiteTested before it's called done.
42 tests in FeatDesk, 48 in FeatReader and 4 for FeatRoom's chain, plus a smoke test of each installed app. Open file formats: if Featbox disappeared tomorrow, your notes would still open.
Under the hood, for technical readers
- Apps: Electron, one process per app, each with its own data, single-instance lock and tray icon. Per-user install, no admin rights.
- Isolation: contextIsolation on, nodeIntegration off, sandbox on, CSP
default-src 'none'. The UI never touches the disk: everything goes through validated IPC. - Network: FeatDesk blocks http(s) and ws in its session; links open in the system browser. No feature uses cloud AI in this version.
- Decision log:
eventos.jsonl, append-only. Each event stores the SHA-256 of the previous one; the app recomputes the chain and shows whether it is intact. Types: request, approval, rejection, message, execution. - Several AIs: the log is an open text file inside the project folder. Any assistant working on the project reads the same context and logs in the same format; no provider gets the whole project.
- Module cards: one short card per module, generated from the code and tied to a commit. If an execution event lands after it, the badge turns "outdated".
- Local model: optional, off by default, localhost only (an OpenAI-compatible server such as LM Studio).
- Tests: 42 (FeatDesk), 48 (FeatReader), 4 for the chain (FeatRoom) and a smoke test of each installed app.
[06] Current status
What exists and what doesn't yet.
No promises beyond what is standing. Version 0.5.1 for Windows, per-user install, no admin rights. When something changes, it changes here.
FeatReaderReader with notebooks, highlights, notes and text-to-speech · free0.5.1 · evaluation
FeatDeskMarkdown notes in a folder you own, no network · paid suite0.5.1 · evaluation
FeatRoomVerifiable decision chain and module cards · paid suite0.5.1 · evaluation
Local memory and "Ask the project"Search your project with a model running on your PCIn development
Pricing and final licenseEverything is unlocked during the pilot; pricing is set with the people who use itAfter the pilot
[07] Who builds it
Who builds it knows operations from the inside.
Renato Manduca, from Garopaba, Brazil. Over a decade inside operations, from the counter to running whole teams: floor, bar, kitchen, staff and cash. That is where the standard for the software he builds today came from: he defines what needs to exist, directs the AI agents that build it, and only calls done what passed the tests.
[08] Pilot
Looking for 5 people for 2 weeks.
I want to know whether Featbox works for anyone besides me. The pilot answers that better than anything on this page.
Who it's for
- You run a long project: a book, a product, research, a business.
- You use more than one AI assistant, or want to.
- You have a PC with Windows 10 or 11.
What you get
- All three apps, fully unlocked during the pilot.
- 20 minutes at the start to install it together.
- Direct help from me when you get stuck.
What I ask
- Use it on your real project for 2 weeks.
- Tell me what got in the way, bluntly.
- A 20-minute conversation at the end.
Your details stay in the email, only to arrange the pilot. Nothing is sold or shared.