Why Nio: it finishes the job, checks its own work and remembers it

Most AI coding tools write code for you. Nio delivers working code: it builds, runs and checks the project, fixes what fails, and remembers the project the next time you come back.

70 seconds, a real recorded run (Nio v0.1.221), not a mockup. Day 1: Nio builds TaskFlow from scratch in 76 s with 8 tests passing. Day 2: "where did we leave off?" is answered from project memory with 0 provider requests and 0 tokens. Nio then adds priorities and due dates in 55 s, with 14 tests passing.

The numbers: Nio + LLM vs direct LLM

Nio isn't a model. It works with your model, and every comparison below uses the same model both ways: once through Nio, once called directly. Every figure comes from a benchmark we ran ourselves and can rerun, with its caveats listed.

1. Nio + LLM finishes what a direct LLM leaves half-done

Calculator app benchmark (2026-10-19, Nio v0.1.249, 8 models, 23 runs)

Nio + LLM spent more in total, but because it finished far more often, each working app cost 58% less. It is the third run of this benchmark (October 5: 7 of 12 vs 1 of 12), and Sonnet 5.5 has scored 100 through Nio in all nine runs. Results and method.

Fix a broken site (2026-10-12, Claude Sonnet 4.6 on both sides)

For about the same money, one side delivered a fixed site and the other didn't.

2. It keeps getting cheaper

We measure cost per finished task, because a cheap answer that doesn't work isn't cheap.

That is half the cost on the same tasks, with 6 of 6 finished, each after an independent review and a passing build. A new app now costs about $0.11. Results and method.

3. Your project's memory costs nothing to recall

• Asking "where did we leave off?" the next day is answered locally: 0 requests, 0 tokens, $0.
• Work that was already verified is reused instead of redone: $0 and about 1 s in 4 of 11 repeat runs.

4. It handles real-size projects

Measured on 2026-10-09 against generated projects:

Why it works

Where Nio is not the best choice (yet)

We'd rather you trust us than be surprised. Measured on Nio v0.1.237 (October 16, 2026), the v0.1.251 changes (October 20) and v0.1.252 (October 21), 3 runs of each task:

• Very small edits cost more through Nio than a single direct model call, because Nio also runs the tests and checks the result. In our v0.1.237 test, a one-line bug fix with Claude Sonnet 5.5 cost about $0.11 through Nio vs $0.05 direct; both fixed it every time. A one-line change doesn't need that.
• Small models gain little: with the v0.1.251 changes, GPT-5 mini averages 98 through Nio and 98 called directly on the calculator benchmark, so the same result for about twice the cost ($0.19 vs $0.09 per run). Claude Haiku 4.5 does gain: 96 through Nio vs 48 direct.
• GPT-5.1 costs more through Nio: since v0.1.252 it averages 98 through Nio vs 91 called directly, both with reasoning on, but at about $1.79 vs $0.64 per run. October 21 results.
• Nio is slower than a single model reply, because it actually runs your builds and tests: that one-line fix took about 30 s through Nio vs 6 s direct. You're paying for a verified result, not for the first draft.

Nio pays off on real work: fixing a broken site, building a feature end to end, coming back to a project the next day.

Try it in two minutes

macOS / Linux

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/luisMan/Nio/releases/latest/download/nio-installer.sh | sh

Windows (PowerShell)

powershell -c "irm https://github.com/luisMan/Nio/releases/latest/download/nio-installer.ps1 | iex"

Then give it something real: a bug you've been putting off, or the app you've been meaning to build. Ask it again tomorrow and see what it remembers.

More videos