i gave my ai agents access to all of social media
media-mcp. an mcp server that lets claude or any mcp client fetch tweets, transcribe videos, download instagram posts, and extract frames from any video url.
i gave my ai agents access to all of social media. 29 tools. 4 platforms today, 21 coming. fully open source.
media-mcp is an mcp server that lets claude or any mcp client fetch tweets, transcribe videos, download instagram posts, and extract frames from any video url. built on @justusecobalt, which already supports 21 platforms.
here is what it can do and why it matters.
the wall
your ai agent lives in a text box. it can read files, write code, browse the web. but ask it to pull a tweet, transcribe a youtube video, or download an instagram carousel and it hits a wall.
social media is one of the richest data sources on the internet and ai agents have zero access to it.
what media-mcp does
29 tools across the platforms i actually use. fetch a tweet by url and get back the text, author, metrics, embedded media. transcribe a youtube video locally with per-token confidence scores. download an instagram post or reel. pull frames from any video url at exact timestamps. search tweets, follow users, monitor accounts, get replies and quotes, check follow relationships.
the foundation is cobalt. one project that already speaks 21 platforms with a shared download protocol. media-mcp is the mcp surface on top, plus the whisper transcription pipeline, plus the per-tool ergonomics for an agent harness.
why it matters
once an agent can read social media, the loop closes. it does not have to ask you to paste a tweet. it does not have to give up when a youtube link is the answer. it can pull the source, transcribe it, extract the frames it needs, and reason about all of it in one pass.
the article on multimodal frugality, ears without eyes, is the architectural sibling to this launch. the cheap channel by default, the expensive one on demand. media-mcp is what makes that pattern run at the wire.