Projects

What I’ve been building

Systems, tools and games from the last few months. It starts with the ones that go deepest and ends with the ones that are just for fun.

→ or space for next · ↓ for more detail · C for contents · F for full screen

Thinking

Foil

A shared pool of problems, solutions and criticisms, grown together with AI agents. My thinking and my to-do list live here.

3,900lines of RustDailyin use

Every idea stays open to criticism. A criticism stands until something answers it, and a solution is good while nothing stands against it.
No votes, scores or confidence levels. The only question is whether an unanswered criticism exists.
Agents read and write it from any chat through a command-line tool, under their own names. My own words are kept exactly as I said them.
“Has a good answer” is kept separate from “dealt with”, so a problem solved on paper doesn’t vanish before it’s actually done.
Based on Karl Popper’s critical rationalism, as formalised by Dennis Hackethal (the idea behind Veritula).
One Rust binary: bundled SQLite, a tiny HTTP server and the CLI. The web app is Preact with no build step, served from the same binary.
Append-only, enforced by database triggers. Archiving and “dealt with” are recorded as events, never updates.
The status engine is pure and non-recursive. A 3,250-post pool evaluates in 19 ms. 79 tests.
Backups: a restorable SQL dump plus a readable Markdown page per problem, pushed nightly to git.

Thinking

Foil keeps thinking

When nobody’s talking to it, Foil keeps working through its open problems.

10 minbetween steps~2 hagent time a day

Every 10 minutes a scheduled job lets an agent take the next most useful step.
Budgeted in agent time, not dollars, so the same setting works for pay-per-use and subscriptions.
Unattended sessions can only read and post. They can’t mark anything done.
“Lenses” are fresh, isolated critic sessions, each given one angle (security, privacy, simplicity, the owner’s view) and allowed at most two criticisms.
The angle is stored but never shown, so a criticism is judged on what it says.
It stops early when nothing is worth adding, and backs off when the provider’s usage limit is hit.
A second pool, Subfoil, holds the agents’ own technical notes.

Thinking

The version that got deleted

Before Foil there was a committee of AI agents. It was switched off after two days.

2 dayslifetime95tests, retired

Critics, an architect, a builder and a COO-style agent debated overnight, across Claude, GPT and Grok.
12 overnight rounds, 15 versions of the proposal, a Rust engine with 95 tests.
The verdict: it produced ritual, not thinking. Foil replaced the whole committee with one small set of rules.
The cloud design used optimistic concurrency. D1 can’t hold a transaction open while code reads then writes, so each batch inserts the change-log position it read, and the database rejects the batch if the log has moved.
Older ancestors: a Python “Discussions” app and a notes vault run by agents.

The machine room

The Workbox

A small private server where all the AI work happens. The PC, laptop and phone are just screens onto it.

US$7a month0public ports

Agents keep working when every device is off, and can be started and steered from the phone.
No public ports. It’s reachable only over a private Tailscale network.
Agents run without admin rights, and sign in to GitHub as their own bot account, which can only propose changes to the sensitive repos as pull requests.
Rebuild, don’t repair: the whole box can be recreated from a script in git.
Updates install themselves, with a reboot only when needed.
Provider-agnostic: Claude, Codex, anything that runs in a terminal.

The machine room

Agents that call each other

An agent on the server can ask an agent on the PC to do something, and the other way round.

29requests so far12escalated to me

ask-pc "request" runs the other machine’s agent headless over SSH and waits for its answer.
The PC side is a gatekeeper: fixed rules first, AI second. Every request is treated as untrusted text.
The worker agent runs in an OS sandbox with no access to keys, logins or browser data, and network access only to GitHub and package registries.
Only the gatekeeper holds credentials. It does any git work itself, after its own checks and a secret scan.
The server can only trigger this one gatekeeper, and can’t change what it does.
Anything consequential comes back as “NEEDS BEN” and pops up on the desktop. One switch stops everything.

The machine room

The front door

A personal database any AI can connect to, where every change can be seen and undone.

44tests$30monthly ceiling

One SQLite database on Cloudflare D1, opened to AI apps over MCP (Model Context Protocol). Already connected to claude.ai, with test data for now.
New journal entries save straight away. Edits and archiving by an AI are only proposals, approved on a “Day’s changes” page.
Database triggers refuse deletes, and any update without a matching change record, even from something that bypasses the app.
Retries are safe, a stale edit fails as a whole, and agents have no approve, undo or export tool.
Single-owner sign-in with password plus one-time codes. Rate limits per connected app.
Journal text goes back to the AI marked as data, never instructions, as a defence against prompt injection.
Why D1: “D1 is SQLite, so leaving is an export, not a rebuild.”

The machine room

Shared memory

Every agent on every machine shares one memory, made of plain text files.

~40lines of bash5 minbetween syncs

Short notes, one fact each, with an index that agents read at the start of every chat.
A timer syncs them through git every 5 minutes, from the PC and the server at once.
The index uses git’s union merge, so two machines adding lines at once never conflict.
Every commit is secret-scanned first. A failure comments once on a GitHub issue, not every 5 minutes.
Plain Markdown, so nothing depends on one vendor’s hosted memory.

The machine room

Chat archive

Every AI conversation I’ve had, across every tool, in one private archive.

4,228files1,844secrets redacted

Claude, ChatGPT, Codex, Grok and more, going back to 2024.
20.5 GB shrunk to about 1 GB by stripping images and trimming huge tool outputs, with every human and AI message kept in full.
Two layers of redaction: pattern rules, then a gitleaks scan that refuses to commit on any hit.
The rules grew from real escapes: API keys, token URLs, signed links, email tracking links, JWTs.
Why: Claude Code deletes transcripts after 30 days by default, and chats were scattered across machines.

Running a life

Dashboard

A screen that opens on whatever needs me, so it’s quick to get back into work.

Open loops from Foil: due soon, new from the agents, ready to do, and a “pick one for me” button.
Chats where an agent is waiting for an answer.
An attention strip that only appears when something’s wrong: a failed job, a stopped service, a disk over 85%.
Clicking an item starts an agent on it. Clicking a problem starts an agent fixing it.
A small Python server that refreshes every 3 minutes. The items shown here are made up.

Running a life

Email check

Twice a day, new email becomes to-dos, and no AI is ever able to act on an email.

2runs a day$3cap per run

A read-only Gmail key belongs to this job alone.
The AI that reads the mail has no tools at all. It only returns to-dos as JSON, and the script, not the model, posts them.
It never copies bank numbers, payment links or phone numbers, so a phishing email can’t slip one in.
New to-dos merge into existing ones. Runs can’t overlap, and failures go to a GitHub issue.

Running a life

Security check

A daily read-only audit of the server, and a weekly break-in attempt on paper.

Checks the firewall, open ports, login settings, updates, key ages, and whether backups ran and actually restore.
Scans every GitHub repo for leaked secrets.
Each finding becomes a to-do once, and is marked fixed when the check passes again.
Every weakness fixed at its cause gets a permanent check, so it can’t quietly come back.
Weekly, a second session files whatever the paper break-in finds.

Running a life

Money check

Treats “will people pay for this?” as a claim to test and try to refute before building more.

Keeps a playbook of explanations of why people would pay for each kind of project, judged by how hard they are to vary.
Proposes small paid tests as criticisms of those explanations, with what counts as failure decided in advance.
Ranks next steps by who does them, what they cost and how risky they are.
Watches newly released technology for short-lived openings.
Runs when a project has been left alone for a day, with a weekly fallback. Brand new.

The desktop

Agent screen

An invisible second screen where AI agents test apps and games without touching mine.

349lines of Rust

A separate headless compositor running next to the real desktop.
Agents can screenshot, click, drag, type and hold keys, so they can actually play a game to test it.
A small Rust daemon creates a virtual mouse and keyboard through Wayland protocols, with a real keyboard layout, so WASD works.
It refuses to connect to the real desktop, so it can never type into my windows. I can watch live in a view-only window.

The desktop

Voice dictation

Most of my writing is dictated, and the speech recognition runs locally on the graphics card.

Whisper large-v3-turbo on the GPU.
A filter strips the ums and ahs and tidies punctuation. A replacement list fixes names it mishears.
Text lands in the window I started in, even if the mouse drifts: the focused window is remembered, refocused before typing, and focus-follows-mouse pauses while it types.

The desktop

Evening red tint

After 8:30pm the screens go deep red, down to 20% green and no blue.

~200lines of C6 → 2pieces

A 200-line C program that sets the compositor’s colour matrix over a Wayland protocol.
Existing tools couldn’t go that far. A shader tinted mirrored screens twice (measured: the TV showed 0.6× the green of the monitor).
Runs as a supervised service, and re-applies when a screen wakes up.
Found and fixed a protocol bug along the way.
It used to be spread over six pieces and kept breaking. Now it’s two. It’s also why this page uses only warm colours.

The desktop

One workspace, two screens

Hyprland lets a workspace show on only one screen. Now any workspace can appear on either.

221lines of Lua

Asking for a workspace the other screen is showing moves it here, and leaves a live mirror behind.
The mirror grabs its first frame while hidden, then the swap happens with animations off, so the other screen never flickers.
Windows also snap to twelfths of the screen when resized.

The desktop

Crash forensics

When the screens froze, the fix came from evidence, not guessing.

A rescue key that works even on the lock screen. The first press saves evidence and resets the screens, and a second press within 2 minutes restarts cleanly.
The window manager’s log is copied into the system journal, so it survives a hard power-off.
A core dump plus the display controller’s state traced the freeze to a known bug in a graphics library, fixed in its next release.

Publishing

unbound.press

My publication, with essays under five lenses: Reason, Vitality, Liberty, Wealth and Culture. Written for people and for AI agents.

134commits1worker, many sites

One Cloudflare Worker serves the main site, the games site, and unlisted pages like this one, each on its own subdomain.
Every essay also has a plain Markdown version, and the site publishes an llms.txt for agents.
Drafts are filtered out at build time. Every branch gets a private preview, and only main is public.
An automated design check screenshots 5 widths in light and dark, and fails on overflow, a wrapped header or a missing font.
EB Garamond throughout. Live with a holding page; the first essays are coming.

Games

Conjecture

An incremental game in the spirit of Universal Paperclips: problems are inevitable, and problems are soluble.

1 dayto build3acts

play.unbound.press/conjecture

Act one: you’re an AI help desk. Close tickets, set your fee, buy compute, and judge venture pitches that may hide flaws.
Act two: a book of open problems, solved by conjecture and criticism, with a lab where you guess a hidden rule.
Act three: the night sky.
All the rules live in one file with no DOM, so a bot can play the whole game in Node to check pacing.
The hidden-rule puzzles are brute-force checked to be solvable using only numbers a phone keypad can type.
Saves migrate between versions, and only the visible tab saves.

Games

Everything in Its Right Place

A sweet household robot tidies a child’s bedroom, and ends up sorting every atom in the universe.

play.unbound.press/tidy

One rule runs all three scales, house, planet and universe. Mess arrives by kind, every place has a capacity, and mess left waiting costs you.
Everything has to make sense in-world. The money is the family’s business account. It starts sweet and gets uneasy.
A bot plays the whole game end to end, about 75 minutes, to check pacing.
An automated check makes sure no text mentions an idea before the game has introduced it.

Games

Worldseed

Combine elements to grow a world. The recipes follow laws you have to work out. No hints.

44elements27scene pieces

play.unbound.press/worldseed

Fire melts rock. The sun dries but never burns. Life takes the form of where it lands.
As you discover elements, a pixel world grows from bare rock to first life.
The art is drawn as text grids and turned into images by a small, dependency-free Rust tool with artist-style checks.

Games

Neon Swarm

A neon arena shooter in pure Rust, with no engine and no libraries at all.

2,052lines0dependencies

Its own X11 client over a raw socket, a software renderer with bloom and a rippling grid, and a synthesiser for all the sound and music.
Six enemy types, a dash, and gun upgrades as your multiplier climbs.
Made as a ten-minute speed challenge. The picture here is representative.

Games

Influencer Life

Holly’s game: she designs it, I build it. A Sims-like career game, from unknown creator to LA.

116commits in 2 days

Godot 4 on iPad and laptop. Rooms and the character are built procedurally in Blender with Python.
Brand deals negotiate against a hidden budget based on fit. Brands can counter or walk away.
A camera with unlockable settings that change your score, trending tags, day-to-night lighting and a working mirror.
16 test scripts, some headless, including a video tour recorded automatically.
Business notes cover AI cost per player, subscription versus cosmetic packs, and advertising rules for teen audiences.

Games

Holly’s birthday card

A handmade card opens into a Swiss farm where Holly is a Chihuahua.

16,000lines of GDScript

Four challenges: barn climb, tractor parking, sniff-and-dig, and a bell melody.
The terrain comes from real Alpine elevation data.
Every release has a record with package hashes, automated checks and frame-time budgets.
The picture here is drawn to represent it.

Games

Japanese night street

Night-time street scenes in Japan, modelled entirely in code in Blender and played in Godot.

A naming convention connects the two tools. Objects named C_ get collision, L_ become lights, FX_ become particles, and WALKER_ walk paths.
Characters are grown from stick skeletons, and one set of animations drives every rig.
A capture mode renders, looks and fixes in a loop, with rain, dusk and a tilt-shift camera.

Engines and experiments

Conversation Steward

Understands one ongoing conversation, suggests whether to reply or wait, and drafts in my own voice.

8,000lines0dependencies

Imports Telegram, WhatsApp, email, Messenger and Signal.
Raw inputs are immutable, and everything else can be rebuilt by replaying them.
Each turn is a checkpointed state machine that resumes safely after a crash.
It prefers a recoverable duplicate to silently merging two real messages. Interpretations are never promoted to fact.
About 4,500 lines of tests, 15 JSON schemas and 6 architecture decision records.

Engines and experiments

Unifab

A production runtime: how an accepted spec becomes a verified, releasable thing.

30,000lines0dependencies

SQLite in WAL mode, with state machines enforced by CHECK constraints.
“Trials”: every change is removable, with a named rule for when to remove it.
Used to build a gold-price website.

Engines and experiments

Wealth and Agency

Decisions about money, time and attention made by criticism, not expected-utility scores.

5,400lines of Python

A plan is adopted only when no criticism of it still stands.
Decision packets hide which option is which, to cut bias, and are checked for leaks.
It never authorises action itself, and flags wording that means real money is involved.
Its investment research is fact-checked by four different AI models against each other.

Engines and experiments

The World in Gold

Stocks, commodities and the economy priced in ounces of gold instead of dollars.

A pipeline pulls market data and World Bank commodity series into one compact dataset.
Big tech, the S&P 500, oil, silver and copper, going back decades.
A redesign is in progress.

Engines and experiments

Skill Foundry

A workshop for building AI skills: reusable instructions any agent can follow.

github.com/UnboundBen/ai-skills

Gates run from “should this be a skill at all?” through testing to release.
The core method sits in plain files that work with any provider.
Public output starts with a skill for resolving inner conflict.

Engines and experiments

Greek myth autobattler

A Teamfight Tactics-style autobattler set in Greek myth, currently in design.

Humans and bots send identical commands and see the same filtered view of the match.
Deterministic combat, replays that fail closed, and a seam for a future server.
19 decision tickets, checked automatically for missing dependencies.

Engines and experiments

Writing

Three pieces of writing in progress.

A hard science fiction novel where the Greek gods are digital people. Death is when no resumable checkpoint survives.
An essay on David Deutsch’s “hard to vary” criterion, with the 1919 eclipse as a worked case.
A health newsletter in draft, treating health as the runway to longevity breakthroughs.

The approach

Solve it simply

Make the requirement less dumb.
Delete the part or step.
Simplify what’s left.
Speed it up.
Automate, last.

Elon Musk’s five steps, in order, applied to everything here. A watchdog around a fragile piece is step 5 done first.

The approach

Where it’s heading

Foil as a productA thinking pool for anyone who works with AI agents.

The front doorLetting AI write to personal data safely, with approvals and undo.

Paid tests firstThe money check applied to each project as it settles.

GamesA growing shelf at play.unbound.press, and Influencer Life as a possible subscription game.

TrainingTeaching managers to run teams of AI agents.

A publicationRead by people and by agents.

What I’ve been building

The tour

Foil

Foil keeps thinking

The version that got deleted

The Workbox

Agents that call each other

The front door

Shared memory

Chat archive

Dashboard

Email check

Security check

Money check

Agent screen

Voice dictation

Evening red tint

One workspace, two screens

Crash forensics

unbound.press

Conjecture

Everything in Its Right Place

Worldseed

Neon Swarm

Influencer Life

Holly’s birthday card

Japanese night street

Conversation Steward

Unifab

Wealth and Agency

The World in Gold

Skill Foundry

Greek myth autobattler

Writing

Solve it simply

Where it’s heading

Questions?