Instruments for thinking, not machines that think for you: a manifesto for research software

Anyone who writes a thesis today has two problems. The first is old: the work is scattered across a reference manager, a word processor, a folder of interviews and a spreadsheet that nobody else can read. The second is new. A text generator offers to write the chapter for you, and the question every examiner asks, "how did you get here?", gets harder to answer with each paragraph you accept.

We build software for the first problem and we refuse the shortcut that creates the second. This note says why, and what that means in practice.

Where the name comes from

Heuretika comes from heuristic, and heuristic from the Greek heuriskein, "to find". It is the art of discovery. Not the art of receiving an answer: the art of looking until something turns up, and then being able to say where you looked.

That is the whole position. We make instruments for thinking and researching. We do not make machines that think for you.

Two temptations

Research has always had two ways of going wrong, and both are cheaper than ever.

The first is what we call the collector's fallacy: confusing the accumulation of material with knowledge. A library of 3,000 PDFs, a notes app with ten thousand cards, a transcript archive nobody has coded. Having it all in one place feels like progress. It is not, until someone reads, compares and decides what it means.

The second is mass generation. A model can produce a literature review in a minute, fluent and plausible. Fluency is the problem. A text that reads well and that you cannot trace back to a source, a decision or a coding choice is a liability when you have to answer for every sentence in front of a committee, a journal or an ethics board.

Between those two poles there is a craft: reading, comparing, coding, writing, and being able to explain how each conclusion was reached. Software can serve that craft. It cannot replace it.

Your files, your machine

The first thing an instrument owes its user is that it outlives the company that sold it. So our desktop applications are local-first. Your work lives in open formats on your own disk, readable without us, and it survives the program. Academic Studio, for instance, keeps your project as plain Markdown files, with Zotero as the source of citations.

The reason is not nostalgia. Scholars know what happens to proprietary formats: Latour, Mauguin and Teil published their socio-technical graph method in 1992, and the HyperCard software that supported it no longer exists. The method is still taught; the tool is gone. A thesis takes four to six years and a book often more. The files should be there at the end.

Bring your own agent

Many researchers already work with an AI agent in the terminal, such as Claude Code or Codex. We think that is reasonable, provided the agent is treated as an instrument.

So we ship no model and hold no keys. You bring your agent, and you direct it. In Academic Studio it works inside your project through a command line (mds) and stage-by-stage skills, and what it edits arrives as tracked suggestions that you can read, accept or reject. Nothing is silently overwritten. If you prefer no AI at all, the Craft build has no agent features.

This matters for a plain reason. Universities are starting to ask researchers to declare and trace their use of AI. A suggestion that is attributed, visible and reversible is easy to declare. A paragraph that appeared in your draft is not.

We are careful with the wording. We say we ship no model and keep no keys, and that nothing leaves your computer unless you send it. We do not say that our applications never talk to a model: Emergent Societies, which is still in development and in private beta, can run language-model agents through a local Ollama or through your own key, and it will say so.

Method made explicit

Software for research is never neutral, because every tool embodies a method. Most hide it. We try to do the opposite: each instrument starts from a published method, names it, and tells you which decisions it makes for you.

ST-Graph implements the socio-technical graph of Latour, Mauguin and Teil, with its indicators checked against the published figures. Grounded, which is still in design, treats Strauss and Corbin's grounded theory as a data model, so that it can show what the theory is missing instead of only counting codes. Emergent Societies asks agent-based models to be specified first, following the ODD protocol, and to be run with fixed seeds and replicas, reporting results with their dispersion and their n.

None of this is a feature list. It is a rule: a tool for researchers should let a reviewer see the method that produced the result.

Pay once

Software that helps you write a book should not become a monthly bill that ends the day your funding does. Our rule is simple: you pay once for software, and monthly only for services that cost us monthly. If we ever offer sync, it will be optional and it will never be required.

We will publish prices when they are settled. Until then, the waiting list is the only way in, and it carries a founding price for the people on it.

An invitation

If this is how you want to work, you can join the waiting list on the home page. You will get a short welcome note now and another when early access opens, and every email has a one-click unsubscribe. We do not use cookies for visitors, we do not track you across the web and we do not resell the list. The form also asks which tool you write in today, because we would like to learn from the answers.

If you think we have it wrong, write to us. A method that cannot be argued with is not much of a method.