Shown on HN Nov 3, it hashes changelog, deprecation, status and pricing pages from the big AI labs every few hours, has two models validate each diff, and posts it to a feed you can filter by model, agent or company, with an optional digest. The maker says the feed already holds about 400 changes, 21 of them breaking (Show HN).
Shown on HN Nov 3, it keeps every version of the tool descriptions from about 336,000 MCP servers, skills and plugins across 20 public sources and grades each change by a written rule, and npx smallprint check or the VS Code/Cursor extension checks your own setup with no account. Daily change briefs can go to email, Slack, Discord or ntfy, and the free plan keeps 25 pins (Show HN)…
Shown on HN Nov 3 as an open source alternative to Sift and Sardine, one decide call returns approve, step up, review or block with reasons, and Claude Code, Cursor or Codex can connect over MCP to build workflows and run model experiments that mix logistic regression or boosted trees with plain language questions to Finic's Grayson model. Self hosting needs Node.js 22 and a Grayson API key, or you can use the hosted Kairos Cloud (Show HN)…
Announced Nov 3 and shown on HN, the site rebuilds every manual page from 1,286 Chainguard OS packages nightly as HTML, Markdown, roff and JSON, and an MCP endpoint at https://man.cgr.dev/mcp with five tools lets an agent look up flags for the exact version an image ships instead of guessing from stale training data. It covers only Chainguard OS packages, not Debian or Ubuntu…
Shown on HN Nov 3, it replays each commit on your fork onto a new upstream release in a sandbox, keeping, dropping or agent rewriting it, then hands you a graft/<version branch only if no new test fails; it also carries patch package and pnpm/bun patched dependencies and ships a Claude Code plugin plus an MCP server. Free covers 10 syncs a month on 5 copies, and Premium is $12 a month (Show HN)…
Shown on HN Nov 3 under MIT, the Go proxy sits in front of a gateway with hundreds of tools and exposes only search tools , describe tool and separate read and write call tools, which the authors' benchmark says cuts tool definition tokens 97% at 500 tools while keeping write calls on their own approval. It passes the caller's own token upstream, runs a multilingual embedding model in process with no GPU, and is early development before 1.0 (Show HN)…
Spotted something we missed? Start a thread.