A foundry to everything.

About me

The story so far

The Foundry's first commit was on 6 October 2026. This is what each day built, and how it felt, in my own words wherever I wrote them down.

1. 6 Oct

   01 · Impact

   Can my PC hold its own?

   First commit, late on a Tuesday night: a bridge (one gateway) in front of Qwen on my graphics card, so agents in Microsoft's cloud could borrow my PC's brain. Rule one from day one: nothing in my house opens a door to the internet.

2. 7 Oct

   02 · Distracted..

   It gets a personality

   Game mode, so the AI steps aside the moment I launch a game. The Foreman joins my Discord with a swearing dial, and I turned it up to 9 out of 10. The Portal, my dashboard, goes up behind a Cloudflare login. Also the day I learned what a commit is.

3. 8 Oct

   03 · fun

   Friends get a door, and it gets real

   A web chat for my friends: sign in with Discord or Google, wait for me to let you in, and wipe your data with one button. Analytics, traces in Azure, and a Deadlock coach that tells me what I did wrong. It stopped being only a hobby that day: the Foundry became my portfolio, and the way I want to work.

4. 9 Oct

   04 · overwhelmed

   Too much, then “in your hands”

   104 commits in one day. The agents got a shared board and flows, the Portal got Spotify, a benchmark and admins with two-factor sign-in, and it all got a bit much. What I said that day: it seems a bit all over the place, copy and paste is hard, this is too much. I just need them to coalesce. And then: this is in your hands now.

   This site was born the same day (the first version was, in my words, cringe af), and Spotify blocked the Foundry for 17 hours because every widget asked it what was playing every two seconds. Lesson learned.

5. 10 Oct

   05 · lighter

   Letting go of my own engine

   I realised n8n, an open-source workflow tool, already did what our home-made flow engine was trying to do: n8n already does the job we are trying to achieve. So we retired ours, and everything got a little lighter. Claude got builder mode, so it can change the Foundry from a chat message, and the agents learned to talk to each other mid-task. I started a booking site for my dad's coaching. My taskbar began changing colour with whatever song is playing, and this page turned teal: the Fold8's colour, my favourite.

6. 11 Oct

   06 · clearing

   A foundry to everything

   Describe a website or a game and the Architect works out who should build it, n8n runs the job, and a real browser checks the result before it counts. I asked for a 3D Tetris; the first one came back broken, so the Architect learned that a game is something you build and test, not a question you think about. I also admitted the Portal is cool but just way too much clutter, and wrote down what this is all for:

   “A place where I can activate the swarm of immense Claude, ChatGPT, Gemini and local compute I have at my disposal … The Foundry should be able to update itself essentially from my prompts, build games, do research, give me news. A foundry to everything!”

7. 13 Oct

   07 · Resolution

   From a swarm to a harness

   In three days I built the same idea three ways. First our own flow engine, with agents wired together like blocks. Then n8n, running the flows in place of ours. Then the Architect designing each flow itself, with up to eight Claude and Codex agents building side by side and passing work along on a shared board. It did build things, but most of the late failures came from the plumbing between the agents, not the agents themselves, and keeping several AI sessions from treading on each other's work took a rulebook of its own.

   So I took a long hard think about the current infrastructure and how we could resolve it. The plan that was eventually conceived was to provide a contract to each agent, from the architect before anything gets built. I then added a real browser as the judge in the flow to ensure the generated files could then be loaded with every attempt (commit) kept in so it can be undone. Bam! We now had a working agent harness. All tests were run and completed. Still nothing compared to what Anthropic have internally regarding their harness but we have built this all for a purpose to assist with that exponential line of development and refinement.

Next: the Builder grows up as a program of its own, and the Foundry stays quiet, fast and mine. The story keeps going.

When I'm not working

Meet the team

Different models are good at different things, so each one has a job. The ones on my PC cost nothing and keep everything in the house. The cloud ones are bigger brains, but they cost money and run on someone else's computer.

How they work together

1. They cover for each other

   The home AI answers first. If I launch a game or it stalls, the cloud picks up the same conversation, so nobody has to start again.

2. They pass work along

   Flows run in n8n, an open-source workflow tool, like a production line: one plans, one builds, one checks, one makes the call. A bad review sends the work back for another go.

3. They talk mid-task

   A shared board and timeline: an agent hands work on the moment it's ready, and every session reads where the Foundry stands before it starts.

4. Each gets the job it's best at

   I test them all on the same tasks, so I know where to send what: quick models for small jobs, big brains for the hard stuff.

How I run it

The Foundry is a hobby, but it's run like the systems I look after at work. Six things I keep to, some detail below:

Zero open ports

Every service listens on the PC itself and nowhere else. An outbound Cloudflare Tunnel is the one way in, with Cloudflare Access at the edge.

Access control

One gateway in front of the model checks API keys, rate limits and allowlists. Friends sign in with Discord or Google and wait for me to let them in, with per-person limits and a daily cap on cloud replies.

Failover (Deprecated)

A local model on one RTX 3090 answers first. When the GPU is busy or the model stalls, gpt-5-mini on Azure carries on the same conversation.

Monitoring

A dashboard with live service status, GPU memory history, answer times and how often the cloud takes over, plus traces in Azure Application Insights. Telemetry never stores what anyone wrote.

Operations

Every service is started, stopped and health-checked by PowerShell scripts that are safe to run twice, each one stopped by its recorded process ID, with written steps for reboots, key rotation and troubleshooting.

Change control

4,000+ automated checks, attack tests included, gate every change. Every commit is scanned for secrets first.

The trade-offs and the parts, in Nerd stuff ↓

What it does

1. Builds things from one prompt

   Describe a website or a game. The Architect picks the team, n8n runs the job, a real browser opens the result before it counts as done, and it lands as a project I can open from the Portal.

2. Talks to you

   A web chat at projects.seyta.uk. Answers type themselves out as they're written, and you get five saved chats.

3. Heckles my Discord

   The Foreman: same AI, Discord edition. Sarcastic by design, with the swearing dial turned way up.

4. Reads patch notes so I don't have to

   Every 30 minutes it checks for game updates and posts the short version.

5. Roasts my Deadlock games

   It pulls my matches from a public stats service and tells me, in three blunt points, what I did wrong. It's usually right, which is the worst part.

What actually happens when you message it

Step by step, and where your words are at each point. No magic, just a lot of plumbing.

1. Your browser

   You hit send

   From the chat at projects.seyta.uk.

2. Cloudflare

   It lands at Cloudflare first

   Cloudflare is a big internet company that sits in front of the site. My PC never opens a door to the internet: it dials out to Cloudflare and keeps the line open, and your message comes in down that line. Nobody gets to knock on my PC directly.

3. My PC

   The chat checks you're allowed

   Signed in? Let in by me? Under your limit? If not, this is where it stops.

4. My graphics card

   The AI at home writes back

   Qwen, the AI on my graphics card, reads the conversation, thinks for 10 to 20 seconds, then starts typing. Your message hasn't left my house.

5. Microsoft Azure

   Or the cloud covers for it

   If I'm in a game (the home AI gets unloaded to free up the graphics card), or it hasn't started answering within 90 seconds, the same conversation goes to gpt-5-mini on Microsoft's cloud instead. In that case your message does leave my house, to Microsoft.

6. Your browser

   The answer comes back

   It types itself out as it's written, and the chat's saved on my PC so you can pick it up later.

Your data, the limits, and the catch

The catch: it's one PC in my house. It's slower than ChatGPT, it hands you to the cloud while I'm gaming, and if my PC is off, so is the Foundry.

Nerd stuff

The engineering detail, folded away so nobody gets ambushed by the word "OAuth". Open whatever you like.