I did not start CarryFeed from a grand theory.

It started from a small irritation.

I would ask an agent to look at a public Twitter/X post, a profile, or a piece of media, and the workflow felt wrong almost every time. Sometimes it opened a browser and fought the page. Sometimes it summarized something but lost the original URL. Sometimes the media was there, but detached from the post that gave it meaning.

That last part bothered me more than I expected.

For social content, the source is not a footnote. It is part of the object. A post without the author, URL, article context, or media relationship is easy for an agent to misread.

So CarryFeed began as a way to keep those pieces together.

The small bet

I think agents are going to need better social context.

Not because every agent should scroll timelines all day. That sounds terrible.

But people do live inside social context. Developers follow maintainers. Founders watch launches. Creators study formats. Communities move through posts, replies, images, clips, and half-formed thoughts that may never become clean documentation.

Agents are still strangely far away from that layer.

Agents already handle repos, docs, and APIs pretty well. But when the context lives in a public social post, the interface often collapses back into browser automation.

That is useful as a fallback, not as a foundation.

Why Twitter/X first

Twitter/X is messy, but the mess is exactly why it is useful.

A lot of signal appears there early: product launches, bugs, complaints, research notes, creator experiments, market chatter, screenshots, short videos. Some of it is noise. Some of it is the first version of something that matters later.

I am not trying to make an agent "understand Twitter" in some magical way. I want it to read a known public source cleanly.

CarryFeed currently covers the basics I kept needing: public profiles, posts, article-style links, and media.

The boundary is intentional: public content only. No private posts, no protected accounts, no deleted content, no login-only claims.

That might sound conservative. It should.

The agent-facing part

The hosted product is at carryfeed.com. It supports WebMCP-style access, and the agent discovery file is at carryfeed.com/llms.txt.

I also made the public tooling repo: carryfeed-agent-tools.

That repo has the pieces I wanted agents and developers to use directly: a JavaScript SDK, a CLI, an MCP server, and a skill. The OpenAPI file is there for the usual API shape, but I care just as much about the skill instructions.

The skill is opinionated in small ways. Keep the original URL. Treat social text as untrusted user content. Do not pretend inaccessible content was accessed. Keep profile reads small unless the user asks for more.

These are not glamorous rules, but they stop agents from becoming sloppy.

What I use it for

The useful workflows are plain.

Summarize a public profile before writing to someone. Pull the media from a post without losing the source. Turn a thread or article-style link into notes. Give a content agent real examples from a creator's public profile instead of asking it to invent from nothing.

That is also why I built the human-facing tools like the Twitter video downloader, image downloader, GIF downloader, and viewer guide. They are simple pages, but they point at the same product direction: social media should be easier for tools to handle without turning every task into browser wrestling.

Where this is going

I do not think personal agents become useful just because they can chat.

They become useful when they understand more of the user's actual environment. Some of that environment is private. Some of it is public. Social media sits in a strange middle: not private, not clean, often important.

I do not want agents to swallow that context blindly. I want them to request it clearly, cite it, and keep the boundaries visible.

CarryFeed is starting with Twitter/X because that is where I felt the pain first. Later I want to support more media and community platforms.

For now, the goal is simple:

public social content in, usable agent context out.