v0.2.0 — Spanish findings, in Spanish

Findings on Spanish text are now written in Spanish. They were not, and the reason is worth stating plainly: the wording came from a string compiled into the engine, so analysing Spanish produced an English sentence wrapped around a Spanish word, and no translator could fix it with a pull request.

Pattern rules never had this problem — they have carried their own message since the beginning. This release applies the same idea to the three analyzers that compute their findings instead of matching a rule, and to the catalog page, which labelled every Spanish word "Overused AI vocabulary."

What changed

• messages in the rule packs. Eight templates covering the overused-word finding, the sentence-rhythm one, the em-dash one, and the catalog's label for lexical rules. A translator can now send them in a pull request, one key at a time.
• RulePack.Messages and RulePack.Text(), plus PackMessages with the defaults and each template's arity. New public API on SignsOfAI.Core.
• The CLI and the MCP server inherit it. Both report through the engine, so a Spanish document now reads Spanish end to end.

Nothing existing changes. A pack with no messages block — every custom catalog already sitting in someone's browser — falls back to English defaults that are byte-identical to the previous output, and English analysis is unchanged. Six tests guard the packs by name: every key present, no key nobody reads, placeholders matching their arity, and Spanish findings free of English.

Also in this release, outside NuGet

• A desktop app for Windows, sharing one interface with the web build. It reads PDF, ODT, EPUB and RTF, reaches a local Ollama with no CORS workaround, scans a whole folder of documents, and measures predictability with a model running inside the app — offline, with nothing uploaded. Download it from the desktop-v0.1.0 release.
• Two new libraries in the repository, not on NuGet yet: SignsOfAI.Onnx (the perplexity engine as an in-process library) and SignsOfAI.Documents (document text extraction).

Versions

SignsOfAI.Core, SignsOfAI.Cli and SignsOfAI.Mcp all move to 0.2.0. A minor rather than a patch: there is new public API and the output a reader sees changed. It also settles a skew — Mcp sat at 0.1.1 from the release that claimed the MCP registry name, while the other two stayed at 0.1.0.

Desktop app 0.1.0 — Signs of AI Writing for Windows

The desktop build of Signs of AI Writing. Same tool as the web app, in a window,
with nothing to install alongside it.

Download

SignsOfAI-Desktop-0.1.0-win-x64.zip — 64 MB, Windows 10 (1809+) or 11, 64-bit.

Unzip the whole folder somewhere and run SignsOfAI.Desktop.exe. Keep the folder
together: the .exe needs the wwwroot next to it.

sha256 52f15134e4c20ff9e060965db24f82433756a1ad1250f8019e960ccf3abcd67b

Windows will warn you. The executable is not code-signed — a certificate costs
a few hundred dollars a year and this is a free tool — so SmartScreen shows
"Windows protected your PC". Click More info → Run anyway. If you would rather
not, the source is right here and dotnet publish src/SignsOfAI.Desktop builds the
same thing on your own machine.

No .NET install needed; the runtime is bundled. WebView2 is required and ships with
Windows 11 and any Windows 10 with a current Edge, so you almost certainly have it.

What it does

• Flags the tells of AI writing in English and Spanish — overused vocabulary,
  rhetorical crutches, syntactic tics, and a rhythm too even to be human — and
  explains how to fix each one.
• Rewrites them out as you type, on the device. No model, no network, no API key.
  When changing a word would break the sentence, it says so instead of breaking it.
• Originality checking between documents: shared passages, with the evidence.
• Interface in English or Spanish, switchable, independent of the language you analyse.

Nothing you write is uploaded. The analysis is ordinary C# running on your machine —
the same engine that is on NuGet as SignsOfAI.Core.

Honestly: what this version does not do yet

This first build does nothing the web app
cannot already do — it is the same interface in a window. It exists because a desktop
app can go where a browser tab cannot, and those are the next three things:

• Local models through Ollama. A page cannot reach localhost:11434 without you
  reconfiguring Ollama's CORS policy. From here it is an ordinary call.
• Predictability without a server. The perplexity model can run inside the app
  instead of calling one — offline, with nothing leaving the machine. The engine
  already ships as SignsOfAI.Onnx.
• Folders of documents. PDF, DOCX, ODT, EPUB, RTF, TXT and Markdown extraction is
  already in SignsOfAI.Documents; wiring it to a drop target is next.

If you are here from Reddit: thanks for asking for this. Tell me what you actually
need it to do — open an issue and
it will shape what lands next.

v0.1.1 — MCP registry ownership

Patch release for SignsOfAI.Mcp only. SignsOfAI.Core and SignsOfAI.Cli remain at 0.1.0.

The official MCP registry verifies package ownership by requiring the server name to appear as mcp-name: io.github.peopleworks/signs-of-ai inside the NuGet package README. NuGet packages are immutable, so proving that needs a new version.

• Package README carries the mcp-name line, so the registry can verify the entry.
• .mcp/server.json moved to the current 2025-12-11 schema, and its description now fits the registry's 100-character limit.

No functional change to the server or its six tools.

Install: dnx SignsOfAI.Mcp --yes · dotnet tool install --global SignsOfAI.Mcp

v0.1.0 — first public release

First public release. Three packages, all MIT.

Package	What it is	Install
SignsOfAI.Core	The engine: rule packs, analyzers, scorer, originality comparison. Pure .NET, no browser.	dotnet add package SignsOfAI.Core
SignsOfAI.Cli	Check drafts from the terminal, or gate prose in CI with --max-score.	dotnet tool install --global SignsOfAI.Cli
SignsOfAI.Mcp	MCP server — the same engine as tools for Claude Desktop and any MCP client.	dnx SignsOfAI.Mcp --yes

What it does

Most AI detectors hand you a number. "87% AI" is unfalsifiable: a teacher can't accuse on it, a student can't appeal it, and a writer can't learn from it.

This one shows the evidence instead. It flags the specific tells — overused vocabulary, rhetorical crutches, syntactic patterns, and the statistical rhythm of sentence lengths (burstiness) — and for each one it says what to do about it. It also compares documents against each other and surfaces the passages they share, including reworded paraphrases across languages.

English and Spanish. The Spanish rule pack is derived from scratch, not machine-translated.

It runs on your machine. The browser app is Blazor WebAssembly; nothing is uploaded, no account. Two optional features (perplexity and cross-language paraphrase) call a server and say so before they do.

Also in this release

• Web app: https://peopleworks.github.io/SignsofAI/
• /signs-of-ai agent skill for Claude Code and other harnesses
• 75 tests, gating this release

It is a heuristic linter, not a verdict machine. That is deliberate.