Windows Agent / First-party evidence
Inspect the project.Not just the promise.
Explore projects created through Donatello Agent and a separately labeled local-model test. Each case separates the original output, our checks and what remains unverified. This is Donatello's own demonstration collection, not an independent review or a general model benchmark.
Microsoft Store · 0.9.87 · 2026-09-24
An actual HTML, CSS and JavaScript website generated in the Windows Agent, with a responsive menu and a clearly labeled demonstration form.
0 Donatello credits · OpenCode Free · space-bunny-free
Inspect files, prompt and checks →Microsoft Store · 0.9.87 · 2026-09-24
Play the actual 4×4 browser game generated by Agent. Inspect the original code, keyboard controls, counters and recorded checks.
0 Donatello credits · OpenCode Free · space-bunny-free
Inspect files, prompt and checks →Microsoft Store · 0.9.87 · 2026-09-24
A controlled JavaScript exercise: a supplied cart-total function ignores quantity. Inspect Agent's repair and the reproducible regression tests.
0 Donatello credits · OpenCode Free · space-bunny-free
Inspect files, prompt and checks →Microsoft Store · 0.9.87 · 2026-09-24
Agent wrote a procedural scene and ran Blender 4.5 locally. Inspect the PNG render, original Python script and editable .blend file.
0 Donatello credits · OpenCode Free · space-bunny-free
Inspect files, prompt and checks →Isolated local-runtime harness · 0.9.87 · 2026-09-24
A controlled run of Donatello's local model: both attempts, a targeted correction, 12 regression checks, actual timings and hardware. Not an automatic-route UI test.
0 Donatello credits · IBM Granite 4.1 3B Instruct · Q4_K_M
Inspect files, prompt and checks →How we document a case
Desktop examples use a project containing only demonstration material. The local-model example instead uses an isolated runtime harness and does not test desktop chat routing. We record requests and observed versions, then preserve original artifacts with byte counts and SHA-256 checksums. We disclose follow-up prompts, manual editing and incomplete checks. Operator checks and presentation wrappers are labeled separately from Agent-generated output.
A file hash verifies published bytes, not correctness or safety for every use. No hosting, reliability, speed or offline guarantee can be inferred from one example. Review generated code before production use.
Windows projects versus web media
These cases concern the Windows coding application and its local-model runtime. The verified media gallery covers the web generators, including Full Video; it is not evidence that the Windows Agent completed those workflows.
Agent overview · Agent Product Facts · Getting started · Security