It is the product where people would earn money rather than only spend attention — the part that makes it more than a network, and the part that makes it slower to ship.

Its top tier carries a free first-tier EchoPage subscription. That is the only place two ORKCOM products are wired to each other, and it runs one way: EchoPage does not depend on YuxHat existing.

The limit

There is no date, and there will not be one until it is a date rather than an invention. It also cannot take a payment until the company is a registered entity, which nobody has finished — said here rather than discovered at a checkout.

Expected October 2026. That is a date, not a promise: when it moves it moves on the news page, and if it slips the page says so rather than quietly staying blank.

What it does is not described here yet. An unbuilt product described in full is a specification handed to whoever reads faster, and there is nothing to weigh against that until there is something to use.

There is no signup and no waitlist to join.

Cogitorium — available, v3.3.0, Apache-2.0

A single Go binary with the whole interface embedded in it. Point it at a frontier API, at Ollama on your laptop, at a box in your homelab, or at all three at once — it does not care, and it does not phone home. Read it

The work has a name: graph engineering. Prompt engineering asks what to say to one model; this asks what the parts may do to each other. Agents are nodes, and four kinds of edge are each a permission the runtime checks — delegation is may-hand-work-to, tools is may-call, memory is what-it-knows-going-in, outward is may-ask-to-search. Remove an edge and the thing it allowed stops being possible.

ContextVerse — available, v1.0.0, BUSL-1.1

You explain yourself to an AI, then switch tools and explain again. Claude has its projects, ChatGPT has its memory, Cursor has its rules; they do not talk to each other, and switching tools starts you from zero. Read it

contextd is one binary holding that space and delivering it into whatever entry point each AI tool actually reads. It runs three ways — alone on your laptop, as a client of a team server, or as the server — and the mode is a question the setup wizard asks rather than a different product you have to buy. Nothing is inferred, embedded or stored behind your back: you curate, the model reads.

dither-ork — available, free, AGPL-3.0-or-later

73 effects, wired rather than stacked: 15 error diffusion, 6 ordered, 13 pattern and halftone, 16 glitch, 17 special, 6 preprocess. A document is nodes and edges rather than a list, so two branches can converge on one node, and a connection that cannot work is refused with the reason before you commit it. Open it

Three effects take no picture at all — noise, gradients and shapes draw from their own parameters, so a document can start from nothing. It runs entirely in the tab, uploads nothing, and needs WebGPU; without one it opens on a screen saying so rather than on a slower version.

grumbo — announced, no date

Announced, not built. It is named here because it is named in the bar, and a name in the bar has to lead somewhere. Early enough that a date would be invented rather than estimated, so there is not one. No app, no store listing, no waitlist. Read it

Dated entries in the company's own words — including the measurements that came back against us, which go in the changelog rather than in a drawer.

It is also where a date moves. EchoPage is expected in October 2026; if that slips, it slips here rather than quietly staying blank on the product page.

The company is not registered, there is no payroll, and hiring someone into that would be unfair to them.

What does exist is a codebase you can read and contribute to, and a template repository where adding support for a new AI client is a pull request rather than a feature request.

Teach it a new tool

Client integrations are templates. If you can name the file your editor reads context from and how it reads it, you can write one.

Break it in public

The bug template asks whether your space was created by the version you are running or carried across an upgrade. That question exists because it found most of our bugs.

Say hello early

There is no careers inbox, because there is no company to route it to yet.

A single Go binary with the whole interface embedded in it. Point it at a frontier API, at Ollama on your laptop, at a box in your homelab, or at all three at once — it does not care, and it does not phone home.

What it is for is workflows that have models in them and still behave like workflows: as deterministic as the parts allow, the same from one run to the next, and reshapeable without being rebuilt. None of that makes a model deterministic. It makes the workflow around one deterministic, which is a different and achievable thing.

Graph engineering

Prompt engineering asks what to say to one model. This asks what the parts may do to each other. Agents are nodes, and four kinds of edge are each a permission the runtime checks — delegation is may-hand-work-to, tools is may-call, memory is what-it-knows-going-in, outward is may-ask-to-search. Remove an edge and the thing it allowed stops being possible.

A fifth layer is drawn on the same canvas and is deliberately not a permission: time is the clocks — what starts on its own, when it next fires and how it went last time. It is on by default, because a workspace where something fires at 03:00 otherwise looks exactly like one where nothing does.

The limit

It is a workbench, not a platform. If what you want is somebody else running the agents for you, this is the wrong tool.

Claude has its projects, ChatGPT has its memory, Cursor has its rules. They do not talk to each other, they do not sync, and switching tools starts you from zero.

For a team it is worse: the thing worth handing over is rarely a fact — it is which script to run, which one never to run, and what to do when it fails. Today that gets re-explained to every AI, every session, by every developer.

One binary, three modes

contextd holds that space and delivers it into whatever entry point each AI tool actually reads. It runs alone on your laptop, as a client of a team server, or as the server — and the mode is a question the setup wizard asks rather than a different product you have to buy.

Nothing is inferred, embedded or stored behind your back. You curate; the model reads. Because stale context is worse than none, a file declares who owns it rather than leaving freshness to somebody remembering.

The limit

BUSL-1.1 is not open source today. Every release converts to a fully open licence four years after it ships, which is written into the licence rather than promised in a blog post.

73 effects, wired rather than stacked: 15 error diffusion, 6 ordered, 13 pattern and halftone, 16 glitch, 17 special, 6 preprocess — any of them any number of times, each with its own opacity and blend mode.

A document is nodes and edges rather than a list, so two branches can converge on one node, and a connection that cannot work is refused with the reason before you commit it.

Three effects take no picture at all. Noise, gradients and shapes draw from their own parameters, so a document can start from nothing rather than from a photograph.

The licence, plainly

Read it, modify it, fork it, charge for what you build on it. What it asks in return is copyleft. The Affero clause reaches one case: run a modified copy somewhere other people use it over a network, and those people are entitled to that copy's source. Opening this page triggers none of that — there is no server, and the copy running in your tab is the copy that is published.

The limit

It needs a browser with WebGPU. Without one it opens on a screen saying so, rather than on a slower version that would misrepresent what it is.

It is named here because it is named in the bar, and a name in the bar has to lead somewhere.

Early enough that a date would be invented rather than estimated, so there is not one. When there is, it goes on the news page.

What it does is not described here yet. An unbuilt product described in full is a specification handed to whoever reads faster, and there is nothing to weigh against that until there is something to use.

There is no app, no store listing and no waitlist to join.