Byakugan: Private, open-source AI-text checker
About
A private, open-source AI-text checker, as a browser app and an MCP server, all on your own machine.
Details
- Author
- JayOfemi
- Downloads
- 346
- Categories
- Other
Jump to
- detect_ai_text returns a 0–100 likelihood with per-signal breakdown and confidence band.
- find_ai_tells identifies exact character spans that read as AI.
- check_reuse performs shingle‑overlap against reference documents you provide.
- check_grammar uses a rule‑based grammar and style checker (Harper) with suggested fixes.
- All tools are deterministic; the server does not rewrite text itself.
- Available as a TypeScript library via npm install @jayofemi/ai-checker.
Setting up with Highlight
This MCP is not yet compatible with Highlight’s one-click setup. However, you can still use it with Highlight by following these steps:
- Download and install Highlight from highlightai.com/download
- Navigate to the plugins tab and select "Add Custom Plugin"
-
Configure the plugin with the settings below
Plugin Name
Byakugan: Private, open-source AI-text checkerCommand (node, npx, python, etc.)Please refer to the README for specific instructions on how to obtain API keys or other required environment variables.
- Enable "Start Automatically" if you want the plugin to start when Highlight launches
From the repository
Add the server to your MCP client’s config with the command npx -y @jayofemi/ai-checker. For Claude Desktop, open Settings → Developer → Edit Config and insert the JSON block. After restarting, ask the assistant in plain language, for example: “use ai-checker to check whether this reads as AI” and paste your text. The server returns analysis; the assistant rewrites from it.
detect_ai_text
Score how likely a passage reads as AI-written, fully on-device, with a per-signal breakdown and a confidence band. Returns a 0 to 100 likelihood, never a yes/no verdict. The signal is directional and biased against non-native English writers, so treat it as a second opinion and not proof. Get a read here, then call find_ai_tells to locate the spans worth rewriting.
find_ai_tells
Locate the exact character spans that read as AI, covering stock phrasing, long dashes, and formulaic sentence openers. Returns a list of spans with start, end, the matched text, and a note. Use these to rephrase only the flagged parts and leave the rest of the writing intact. Deterministic.
check_reuse
Compare the text against reference documents you provide, using word shingles, and return the share of overlap and the verbatim shared passages. This is a local check against the documents passed in. It does not search the web or any plagiarism database. Deterministic.
check_grammar
Run a rule-based grammar and style check over the text, fully on-device, and return each issue with its character span, a message, the flagged text, and suggested fixes. Use it to tidy grammar and phrasing, before or after rewriting the flagged AI tells. Deterministic.
Claude Desktop / Cursor
Paste into your MCP client config file to install this server.
{
"mcpServers": {
"byakugan: private, open-source ai-text checker": {
"ai-checker": {
"command": "npx",
"args": [
"-y",
"@jayofemi/ai-checker"
]
}
}
}
}McpServers
{
"ai-checker": {
"command": "npx",
"args": [
"-y",
"@jayofemi/ai-checker"
]
}
}AI Checker (MCP) 
AI-text checks inside the AI assistant you already use, on your machine. Add it to your MCP client and it gives that assistant a detection read, the exact phrases that read as AI, a reuse check, and a grammar pass, so the assistant can act on them without your text going to a website. Because the rewriting runs on your own model, a stronger model gives better results than the small one the site runs in the browser. Ships as an MCP server and a TypeScript library.
Install
Add it to your MCP client's config. For Claude Desktop, open Settings, then Developer, then Edit Config, and add: ``json
{
"mcpServers": {
"ai-checker": {
"command": "npx",
"args": ["-y", "@jayofemi/ai-checker"]
}
}
}
`
Restart the client. For other MCP clients (IDEs and the like), point them at the command npx -y @jayofemi/ai-checker over stdio.
Use
Ask the assistant in plain language:
- "use ai-checker to check whether this reads as AI" (then paste your text)
- "use ai-checker to find the parts that read as AI, then rewrite only those"
- "use ai-checker to check this draft against my notes for reuse"
- "use ai-checker to grammar-check this draft"
The server returns the analysis; the assistant rewrites from it.
The tools
The server is deterministic. It hands the assistant precise local analysis to work from, and does not rewrite text itself.
- detect_ai_text - a 0 to 100 likelihood that the text reads as AI, with a per-signal breakdown and a confidence band. Never a yes/no verdict.
- find_ai_tells - the exact character spans that read as AI (stock phrasing, long dashes, formulaic openers), so the assistant rephrases only those.
- check_reuse - shingle-overlap against reference documents you pass in, with the shared passages. A local check against what you provide, not a web search.
- check_grammar - a rule-based grammar and style check (Harper), returning each issue with its span, a message, the flagged text, and suggested fixes. Deterministic.
Use as a library
`
npm install @jayofemi/ai-checker
`
`ts
import { detect, findTells, checkOverlap, checkGrammar } from '@jayofemi/ai-checker';
detect("Your text here."); // likelihood, band, confidence, signals, caveats
findTells("Your text here."); // the spans that read as AI
checkOverlap("your draft", ["a reference"]); // shingle overlap
await checkGrammar("Your text here."); // grammar issues with fixes (async)
`
What to expect
Detection is a directional signal, not proof, and it is biased against non-native English writers. Even, formulaic, stock-phrase-heavy writing reads as more AI-like. Treat it as a second opinion and use it on your own writing.
Develop
`
npm install
npm run build
npm test
`
Verify the server with the MCP Inspector:
`
npx @modelcontextprotocol/inspector node dist/server.js
`
It opens a local UI; connect over stdio, list the tools, and call them. Or run the server directly with npm run mcp`.
License
MIT. Copyright (c) 2026 Jay Ofemi.Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
