Open source projects
Three free tools I develop and maintain, born from problems I encounter working on the web. The code is public: anyone can use, inspect and improve it.
- Licence
- MIT
- External dependencies
- None
- npm releases
- Signed
$ npx stagecheck1 staged release waiting for approval.color-utils 2.3.0 → 2.3.1 staged by ci-bot (token) files: 1 added, 1 modified, 0 removed · size 271 B → 685 B (+153%) HIGH package.json adds a "postinstall" script, which runs automatically on install HIGH bundle.js reads credentials or credential files (×2) HIGH bundle.js contacts a service commonly used to exfiltrate data HIGH bundle.js references GitHub Actions workflow files (worms such as Shai-Hulud plant workflows) HIGH bundle.js runs npm account or publish commands from shipped code (self-propagating worms do this) MEDIUM staged by ci-bot (token), but the previous version was published by alice MEDIUM bundle.js starts other programs (×2) low contacts 1 domain not referenced by the previous version 5 high-risk findings: review carefully before approving.color-utils@2.3.1: [a]pprove, [r]eject, [s]kip, [q]uit? (s)
Sample stagecheck output for a fictional compromised release: the color-utils package and the attack are invented.
$ npx crawlerid verify --ua "Mozilla/5.0 (compatible; GPTBot/1.1)" --ip 203.0.113.7FAILED GPTBot (OpenAI, ai-training) [ip-not-in-published-ranges]$ npx crawlerid verify --ua "Mozilla/5.0 (compatible; GPTBot/1.1)" --ip 20.171.207.12verified GPTBot (OpenAI, ai-training) [via ip-range]$ npx crawlerid robots --disallow ai-training# Generated by crawlerid 0.1.1 - https://github.com/daniel63194/crawlerid# ai-training: Google, Apple, OpenAI, Anthropic, Mistral AI, Common Crawl, Amazon, MetaUser-agent: Google-ExtendedUser-agent: Applebot-ExtendedUser-agent: GPTBotUser-agent: ClaudeBotUser-agent: MistralAI-TrainingUser-agent: CCBotUser-agent: AmazonbotUser-agent: meta-externalagentDisallow: /
crawlerid output. 203.0.113.7 is reserved for examples; 20.171.207.12 is in the list published by OpenAI.
$ npx whowrote -C vscode --since 2026-04-01 --by quarterwhowrote · vscode · 12,285 commits · 2026-04-01 → 2026-10-13 No AI disclosed 5,783 47.1% ███████████ AI-assisted 5,771 47.0% ███████████ AI agent 413 3.4% █ Bot 318 2.6% █ Period Commits No AI disclosed AI-assisted AI agent Bot 2026-Q2 5,757 64.3% 30.3% 2.4% 2.9% 2026-Q3 5,669 33.2% 59.9% 4.5% 2.4% 2026-Q4 859 23.1% 73.1% 2.2% 1.6%
Sample whowrote output for VS Code's history, from April to mid-October 2026.
stagecheck (code on GitHub, opens in a new tab)
Is this update safe to publish?
Many npm package attacks arrive as an ordinary-looking update: someone steals a publisher's access, uploads a version with a hidden script, and anyone installing it runs that script unknowingly.
stagecheck compares the new version with the latest published one before it goes live. It flags the changes these attacks rely on: install scripts, credential access, data sent to external services, secretly added GitHub workflows and self-publishing code. With npm's two-stage publishing, the release waits until you approve it.
- Who it is for
- Anyone publishing npm packages, manually or through a CI pipeline
- How to use it
- From the terminal with
npx stagecheck, or as a GitHub Action
crawlerid (code on GitHub, opens in a new tab)
Is it really GPTBot?
Anyone can present themselves to a website as Googlebot or ClaudeBot: they just need to write it in the User-Agent, and many scrapers do this to bypass rules intended for real crawlers.
crawlerid recognises 47 crawlers and AI agents from 12 operators, including Google, Microsoft, OpenAI, Anthropic and Perplexity, and checks they are who they claim to be: it compares their address with each operator's published IP lists and checks reverse DNS, the method recommended by Google and Bing. It also generates robots.txt by category, so you can exclude AI training while remaining a source for ChatGPT and Perplexity web searches.
- Who it is for
- Website, API and firewall administrators
- How to use it
- As a JavaScript library, or from the terminal to find fake crawlers in server logs
- Runs on
- Node.js, Bun, Deno, Cloudflare Workers and Vercel Edge
whowrote (code on GitHub, opens in a new tab)
Who wrote this code?
More and more code is created with AI agents, but project history does not reveal how much. In the VS Code repository, commits disclosing AI use rose from 33% to 64% between the second and third quarters of 2026.
whowrote reads commits and pull requests and classifies them as people, bots or AI agentsbased on what they disclose: co-authors, agent accounts, branch names, session links and attribution notes. It does not guess: when no AI use is disclosed, the result is "no AI disclosed", not "written by a person".
- Who it is for
- Maintainers and teams who want to understand AI use in their projects
- How to use it
- From the terminal, as a JavaScript library, or as a GitHub Action that labels pull requests
- Recognises
- 39 tools: 19 AI agents, 7 AI reviewers and 13 bots
What this means for your website
These problems affect business websites too, and I bring what I learn from these projects into the websites I build.
- You can choose which AI crawlers can read your website: for example, exclude model training while remaining a source for AI-powered searches.
- Bots pretending to be Googlebot can be identified and blocked without affecting the real ones.
- Plugin and package updates are a common route for attacks: I know what to check before installing them.
- The code is public: you can see how I work before trusting me with a project.
Contributions to other projects
When I find a bug in a tool I use, I try to fix it at the source.
-
nuxt-modules/robots
The module that generates robots.txt and crawler rules for Nuxt websites.
-
ossf/criticality_score
The OpenSSF tool that estimates how critical an open source project is to the ecosystem.