EleosTen dead ends between download and first answer
We installed a shipped release on clean Windows machines and used it as a stranger would, screenshot by screenshot. None of this was found by reading code.
Why publish this
Every desktop AI tool has a first five minutes, and almost nobody measures theirs. We did, on wiped Windows 11 installs with no developer tooling, a machine as ordinary as the ones customers use. The findings are specific to our app but the failures are not: most of them are what happens when software written by people with Node, Git and a C++ runtime already installed meets a machine that has none.
Free to cite. Quote any of it with a link back to this page.
Method
- A packaged release, downloaded from the public CDN, not a development build.
- Clean Windows 11 virtual machines, no Microsoft C++ runtime, no Node, no Git.
- Run as a first-time customer: install, pick something, try to get one answer.
- Dated 13 September 2026, on release 0.1.32; fixes shipped in 0.1.32 and 0.1.34.
What stopped people
| # | What the customer hit | What it really was |
|---|---|---|
| 1 | The main button made an account with no model, then the first message returned a raw error | Onboarding asked the wrong question first: it should ask which AI plan you already pay for |
| 2 | Choosing a plan unmounted the screen mid-login and left a terminal running behind the composer | The app decided which screen to show from whether any account existed, instead of from an explicit "still setting up" flag |
| 3 | Going back left a half-made account behind | Abandoning setup removed the profile but not the pending sign-in |
| 4 | "Download a model in Settings → Models" with no way to get there | The hint named a screen the customer could not reach from where they stood |
| 5 | A machine too small for any model was told "cloud models still work" | Addressed to somebody who had just said they have no subscription |
| 6 | Choosing ChatGPT, Cursor or Grok opened a PowerShell window saying npm i -g … | The command-line agent those plans need was missing, and a person on a ChatGPT plan has no reason to own Node |
| 7 | Every plan opened a terminal to sign in | Two of them had an in-app sign-in all along; only Claude needs the console |
| 8 | Even when hidden, one agent popped an empty black console | The no-window flag does not reach a grandchild process; the platform binary has to be run directly |
| 9 | You could download a model and still be told to download a model | Downloading did not set the model as the one to use |
| 10 | A downloaded model would not start: exit code −1073741515 | STATUS_DLL_NOT_FOUND — the model server needs Microsoft's C++ runtime, which a clean Windows does not have |
The one that was not a user-interface problem
After all ten, a local model still never answered. Three causes, each measured on the machine rather than reasoned about:
- Thinking. The agent harness asks for adaptive thinking; a 2B model spent 5,458 hidden tokens on "17 times 23" and never produced a reply. Disabling thinking for local models made the same request answer in 78 seconds.
- The prompt floor. The harness's first request is around 25,200 tokens with every tool enabled, 6,144 with the local set. Model builds published with 4,096 or 8,192 windows could not answer at all, and now are not offered.
- The window the CLI believes in. For an unfamiliar model name, the maximum-context setting is the window. At 32,768 it compacted before the first message, every turn, forever.
The memory side of the same problem — what each build costs to load, and why a context you can afford is not the same as one an agent can work in — is measured in what fits in 8 GB of VRAM.
What this cannot tell you
- One product, one platform. The onboarding failures are ours; the runtime and context findings apply to anything running llama.cpp behind an agent harness.
- Virtual machines, not laptops. Speed numbers from them are not worth quoting.
- No sample of real customers: this is what an attentive tester hit, not a measured funnel. Since 20 September 2026 the app does count one thing at startup — the release number and whether this machine had started it before, with no identifier attached — because without it we could not tell an install nobody opens from one everybody abandons at the same screen. Everything above predates that count and none of it rests on one.
What we changed
Onboarding now asks one question — whose subscription do you already pay for — and everything else follows from the answer. The app installs a missing agent with your consent instead of showing a package-manager command. Two of the four plans sign in without a console at all. The C++ runtime ships with the app and is copied beside the model server. Local models default to thinking off, a sane context floor, and a window the CLI can believe.
If you build something similar: wipe a machine, install your own release, and watch someone use it. Not one of the ten came out of reading the code, and the three runtime causes needed a logging proxy and a real model rather than an argument about what ought to happen.
Try it on Windows. Seven days free, no card, then $19.90/month for two PCs. The AI is whichever plan you already pay for, or an open model on your own PC, which needs no subscription.