Spectrum Labs

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

Spectrum Labs AI

PullMD v2.4.1 is out - claude.ai web custom connector works natively now, plus what 2 weeks of your feedback turned into

Two weeks ago I posted PullMD here. 385 upvotes, around 60 comments, a bit over 20 GitHub issues, and 7 releases (v1.1.3 → v2.4.0) in 14 days. That was a great experience - and this sub in particular has been a genuinely good place to share something. So: thanks! Quick refresher for anyone who missed the first post: PullMD turns any URL into clean Markdown via MCP, fully self-hosted. Three services in Docker (main app + Trafilatura sidecar + optional Playwright sidecar for JS-heavy pages), zero third-party LLM calls, ships an MCP server so Claude Code / Claude Desktop / claude.ai web can pull clean content directly instead of parsing HTML in your context window. This post is what's new and how to get it. What's new claude.ai web + Claude Desktop work natively now This is the biggest unlock from v2.x. The claude.ai web custom-connector dialog and Claude Desktop's custom-connector dialog now both work against self-hosted PullMD instances. So you can point claude.ai at your own homelab box, hit "Add custom connector," and it works end-to-end. Setup is two env vars: OAUTH_JWT_SECRET=$(openssl rand -hex 32) PUBLIC_URL=https://your-host.example.com Restart. Then in claude.ai web → Settings → Connectors → Add custom, point at https://your-host.example.com/mcp. The connector dialog discovers the server's metadata, registers itself, and walks you through a consent screen. Same flow works in Claude Desktop. Under the hood: standard OAuth 2.1 Authorization Code flow with PKCE-S256 and Dynamic Client Registration - RFC-compliant so any spec-compliant MCP client should work, not just claude.ai/Desktop. Opt-in: if OAUTH_JWT_SECRET isn't set, behavior is identical to v1.x. The Anthropic-side claude-ai-mcp#237 proxy bug I flagged in EDIT2 of post 1 has cleared on their end - though in hindsight, a forgotten custom WAF rule on my side was likely the actual culprit anyway. Verified end-to-end against both dialogs. Multi-user auth Until v2.0, PullMD was effectively single-tenant - a personal homelab tool, open like a barn door to anyone who landed on it. v2.0 adds three auth modes via PULLMD_AUTH_MODE: disabled - the default. Identical to v1.x. No login, no API key required. Right if you're the only one using your instance and you trust your network. single-admin - one user, password-protected, no self-signup. Right for a homelab box where you want the GUI gated but don't want to manage users. multi-user - self-signup at /signup, per-user history isolation, per-user API keys. Right for a shared instance (team, office, friend group). API keys are pmd_ , sent as Authorization: Bearer pmd_xxx, managed at /settings. Share links (/s/:id) stay public in all modes - the whole point of a share link is to be shareable. Minimal upgrade for a shared instance: PULLMD_AUTH_MODE=multi-user PULLMD_ADMIN_EMAIL=you@example.com PULLMD_ADMIN_PASSWORD=change-me-please PullMD works on more sites A bunch of things in v1.2 and v2.2 together close gaps where PullMD used to silently return half-articles, empty bodies, or garbled text: Future PLC family (windowscentral.com, tomshardware.com, techradar.com, pcgamer.com, gamesradar.com, t3.com) used to return mangled content because Readability got confused by recommendation widgets stuffed mid-article and an aria-hidden paywall pattern. The default site-recipes shipped with v2.2 strip both, no config needed. GitHub Issues pages used to return only the original issue body - the JS-rendered comment thread never made it in. The default recipe for */*/issues/* now forces Playwright with wait_for: .js-comment-body, so you get the full comment tree. Sites that fingerprinted the old hardcoded Chrome 131 UA now extract cleanly - UA rotation pulls from a real-world UA pool that updates regularly (v1.2). Pages with navigator.webdriver-style anti-bot detection go through more often - the headless-Chromium sidecar bundles playwright-stealth (v2.2). Sites without an explicit charset declaration (a lot of older German news sites, for example) no longer return mojibake - charset is detected from the byte stream when the response is silent (v1.2). If you have a specific site that still misbehaves, v2.2 lets you (or your Claude Code) write your own recipe - declarative JSON with four rule categories (preprocess, fetch, select, extractor). Drop it at data/site-recipes.json and your rules layer on top of the defaults. There's also a /api/recipes/status endpoint for monitoring. Web GUI: rendered Markdown view + persistent settings Two smaller improvements in the browser frontend (the PWA you get when you open your PullMD instance directly): Rendered Markdown toggle. The result header now has a Raw | Rendered switch, so you can read what you pulled as formatted HTML directly in the browser instead of squinting at the source. Raw stays the default; your choice persists across sessions (v2.4). Settings persist across reloads - frontmatter toggle, comments toggle, comment-depth input.