Shown on HN Nov 1, the test installed 67 npm MCP servers in a sandbox with no internet route and no real credentials, called no tools, and found 7 looked up undeclared domains at startup while 24 offered tools their README doesn't name (2 did both). It's one run that only reads the shipped README, so the author says it isn't a ranking, and 29 finding pages stay unpublished until hand-checked.
Shown on HN Nov 1 under MIT/Apache 2.0, parts are plain JSON step programs, and claude mcp add geop geop mcp (or the Codex equivalent) lets your agent sketch, extrude, fillet, inspect and export STEP or STL through the same editor as the browser app. The in browser assistant sends your messages and program to Geop's server and Anthropic, while the CLI and MCP route run locally (Show HN)…
Shown on HN this week, the Rust daemon keeps a live mirror of each screen's accessibility tree and frames, with no APK or XCTest runner needed on Android or the Simulator, and its CLI copies agent browser's command grammar while mobile agent mcp serves the same tools to Claude Code, Codex or Cursor. It's an experimental pre release, built from source only (macOS for everything, Linux for Android), and leans on private Apple and hidden Android APIs (Show HN)…
Shown on HN Nov 1 under MIT, the Docker stack polls OpenProject 17 work packages, dispatches the pi coding agent as frontend, backend or designer to open a PR, then hands it to a reviewer agent that merges it or kicks it back for fixes, with the engine (never the agents) doing every status change. It needs one time OpenProject admin setup, a GitHub token that can push and merge, and one model provider key such as OpenRouter or a local llama.cpp (Show HN)…
Shown on HN Nov 1, relayvora handoff from codex to claude verify auto confirm reads your latest session, captures diffs and commits, runs the project's own tests, and launches the other agent with a briefing that labels each fact as observed, user, agent reported or inferred. It's local only and pre alpha, and the author says it saves context on big sessions, not money on small ones…
Shown on HN Nov 1 under Apache 2.0, npm install g sweet search , then sweet search init and sweet search index , gives your agent six fully local search, grep, trace and read tools plus a tuned prompt, with the index updating itself as you edit. On the author's own 200 task SWE rebench run it cut agent cost 5–33% with the same number of tasks solved, but it's macOS and Linux only (Windows via WSL2) and costs GPT models about 2.5 points of retrieval accuracy…
Spotted something we missed? Start a thread.