aireviewrating · section inventory

AI agent tools, harvested from a source that only posts repos

Every post on the Instagram account @marc.kaz was read — 1,003 posts back to 2012 — and every GitHub repository it pointed at was pulled out, de-duplicated and verified against the GitHub API. The result is 159 real projects, of which 37 are genuinely AI agent tooling: harnesses, skills, sandboxes, MCP servers and orchestration frameworks.

159repositories, de-duplicated
37filed under AI Agents
148new to the AIRR catalogue
4.47MGitHub stars, combined
153confirmed by the post's own link

How each row was established

The account almost never puts the repository URL in the caption — only 137 of 1,003 captions carry one. It is burned into the video’s text overlay instead. Every post’s cover image was therefore downloaded and read with Apple’s Vision OCR, and the URLs recovered from that text were then looked up against the GitHub API. Nothing here rests on the OCR alone: a row exists only because api.github.com/repos/<owner>/<repo> returned a live repository.

153 rows are linked in overlay — the post’s own on-screen URL resolves to that repository. The remaining rows are named in post: no URL was on screen, but the project is named in the caption and matches a canonical repository. Nine candidates were discarded for being passing mentions or same-name-different-project collisions, and 102 posts carry no repository at all — they are news, robotics clips or paid-product promos, and are correctly absent.

Descriptions, star counts, licences and languages come from GitHub, never from the Instagram captions. 28 repositories publish no licence file or an unrecognised one, which is flagged per row and matters before any of them is listed as usable.