OpenMAIC and the authoring seat you use twice a year

The safety manual at a 90-truck plumbing company is a 46 page PDF that everybody signs on their first morning and nobody opens again. Turning it into real training, the kind with slides and a narrator and a quiz at the end, has two prices. You buy one Articulate 360 Teams seat at $1,749 a year and hand it to whoever on staff seems most likely to figure it out, or you send the manual to a course development shop and pay somewhere in the five figures for a finished hour. OpenMAIC takes the PDF and gives you back the course. It is MIT licensed, it picked up roughly 4,200 stars this week, and it shipped a release yesterday.

What it actually is

OpenMAIC comes out of a lab at Tsinghua, and the idea behind it is older than the repo. The lab's argument, published this year in the Journal of Computer Science and Technology, is that a MOOC is one video played at N students, and that the thing agents make possible is the inversion: N agents aimed at one student. They ran it on more than 700 of their own students across two years before any of this was a product you could install.

What that turns into, practically, is this. You give it a topic, or you upload a document, and it builds a classroom. Not a slide deck. A classroom, with four kinds of scene in it: narrated slides where an agent draws on a whiteboard while it talks, quizzes, interactive HTML simulations for the parts where somebody needs to try the thing rather than watch it, and project work. There is an AI instructor and there are AI classmates, and they speak out loud and interrupt each other, which sounds like a gimmick until you remember that the reason nobody finishes the compliance video is that the compliance video never once addresses them.

It eats PDF, Word, PowerPoint, audio and video. That last part matters more than it reads. Most small companies do not have a training problem, they have a training-is-trapped-in-a-recording problem: there is a two hour Zoom from eighteen months ago where the operations lead explained the whole intake process, and it has been sitting in a shared drive ever since because nobody has four days to turn it into anything.

The August 27 release, v1.0.0, added the part that makes it usable by someone who is not an instructional designer. They call it the agent workbench, and it is a chat window that plans, builds and revises a whole course while you talk to it. Sessions are server backed now, so you can close the laptop and come back. There are twenty built-in skills covering curriculum planning, teaching styles and slide craft, and there is a PowerPoint importer, so the deck you already made is an input rather than a competitor.

Then the exports, which are the reason to take it seriously rather than bookmark it. It writes editable .pptx with the images, charts and equations intact. It writes self-contained interactive HTML. It writes a classroom archive with the external assets inlined so the thing plays with no internet. You are not adopting a format somebody can take away from you later.

Why this lands differently at a 90 person company

The reason training at small companies is a PDF and a signature page is not that owners do not care. It is that every tool in the category is priced and shaped for somebody whose entire job is building courses, and no company under a few hundred people has that person.

Run the seat math honestly. $1,749 a year buys one Articulate license. A distributor or a trades business builds, realistically, two courses a year: the new hire onboarding and whatever the insurer or the state asked for. That is $875 a course before a single hour of anybody's time, for software that sits closed for ten months. And it is not the expensive half. The Chapman Alliance study that industry still quotes found it took something like 79 hours of work to produce one hour of the simplest tier of e-learning, and up into the hundreds for anything interactive. Price that at a competent operations person's loaded rate and the seat is a rounding error.

Outsourcing is the other door and it is worse on the number. The 2010 Chapman benchmark landed near ten thousand dollars for a finished hour of interactive e-learning, and restated for 2026 inflation that band runs from about $14,300 for a basic page turner up past $70,000 for anything with real simulation in it. Nobody with 90 employees is signing that for a forklift refresher. So they do not make the course. They send the PDF.

The practitioner version of the same trap is quieter and just as expensive. An ops or enablement lead at a SaaS company has the new-rep ramp living in three places: a Notion page, a deck from the last kickoff, and a recording of the person who actually knows how discounting works. Turning that into a real ramp is a project, projects need a quarter, and the quarter never comes. What OpenMAIC changes is the size of the first attempt. Feeding it the deck and the recording and getting a narrated course with a quiz back the same afternoon is not the finished thing, but it is a draft you can react to, and reacting to a draft is a completely different task than starting one.

There is a second-order effect worth naming. Once the cost of producing a course drops to an afternoon, courses stop needing to be permanent. The reason onboarding material at most companies is three years stale is that updating it costs almost as much as building it, so nobody does, so it quietly becomes fiction. Cheap regeneration is what makes it honest.

The honest accounting

Free is doing a lot of work in everything above, so here is the rest of it.

Setup is genuinely not hard, and it is also not one click. There is a Docker compose file, there is a one-click Vercel deploy, and there is a hosted demo at open.maic.chat if you only want to see whether the output is any good. If you have a domain and an API key already, you are looking at an afternoon to a running instance. Getting to a course you would actually show a new hire is two or three days, because the first generation of anything is generic and the work is in the editing. That is not a knock on the tool. It is the same two or three days you would have spent in Articulate, minus the two weeks of learning Articulate.

The monthly bill is not zero and it is not flat. OpenMAIC does not sell you inference, it asks for your own provider key and passes the cost through, which means a course costs what a course costs in tokens, and narration is billed per character by most text-to-speech providers. A small VPS at twelve to twenty-five dollars will run the app and a Postgres. The optional MP4 export will not run comfortably there, because rendering video means running Chromium and FFmpeg, and you should budget a bigger box or skip video entirely and ship the interactive HTML, which is better for training anyway.

Now the gap that decides whether this works for you, and it is the one the README tells you if you read the export table rather than the headline. Three formats: PowerPoint, HTML, classroom archive. None of them is SCORM. There is no xAPI. There is no handoff into an LMS, which means there is no completion record.

For onboarding and enablement, who cares. For compliance training, the record is the product. The entire reason a company buys the forklift course is so that after the incident there is a timestamped row proving the person took it. OpenMAIC will make you a better forklift course than your LMS vendor will, and it will not make you the row. If your training exists to satisfy an insurer, a regulator or a customer's vendor questionnaire, this replaces your authoring tool and not your LMS, and you need to be clear-eyed about that before you cancel anything. Articulate also carries accessibility conformance work and a stakeholder review workflow that this does not have, and if you are a government contractor those are not optional extras.

Then there is who maintains it. This is a university lab, which cuts both ways. The MIT license, which they moved to at v0.3.0 back in June, means the code outlives the funding no matter what. The release cadence probably does not. A lab ships because a paper needs it to ship, and when the students who wrote it graduate, the velocity you are looking at this week is not a promise about next year. The contributor list on recent commits is broad enough that this is a real concern rather than an obvious one, but plan as if you are adopting the version you install.

And the part that would be irresponsible to leave out. Seven security advisories have been published on this repository in the last nine days, and two of them are rated critical. Both criticals are the same shape: a media proxy endpoint could be talked into fetching a cloud provider's internal metadata address, which is the address a server asks to find out what credentials it has been issued. In plain terms, an instance sitting on a cloud VM could be tricked into reading out its own keys. There is also an authorization hole that let any visitor overwrite an existing classroom, and a stored cross-site scripting issue in course content.

All of it is patched, in v1.0.1 on September 6 and v1.0.2 yesterday. I do not read a run of advisories like that as a reason to avoid the project. Software that is being looked at hard, and that fixes things in public with a named GHSA and a changelog entry, is in better shape than software nobody has bothered to test. But it is the clearest possible statement of what self-hosting actually costs, and it is not the twenty dollar VPS. Somebody at your company has to watch a releases page. If you deploy this in September and touch it again in March, you are running a version with a known critical hole and a public description of how to use it, and the vendor whose job that would normally be is you now.

Two practical consequences. Turn on the access code before you show anyone, because the default posture is open. And do not put your HR handbook, your incident reports or anything with a name in it into the hosted demo, which is a research lab's convenience service and not a vendor with a data processing agreement.

What is left

The interesting thing here is not that course software got cheaper. It is what the seat was actually buying.

A $1,749 license does not stop anyone from making training. What it does is make training a job, and once it is a job it belongs to a person, and at a 90 person company there is no such person, so it belongs to nobody and the manual stays a PDF. The tool was never the constraint. The constraint was that the tool implied a role.

Every company already owns the course. It is in the manual everybody signed, the deck from the kickoff, and the recording of the one person who actually knows how the thing works. Nobody was ever missing the content.