you-read-i-read

<table align="center" border="0" cellpadding="0" cellspacing="0"> <tr> <td valign="middle" width="200" align="center"> <img src="assets/icon.svg" alt="" width="160"> </td> <td valign="middle" width="540"> <h1>You Read, I Read</h1> <p>A persistent reading agent for <a href="https://claude.com/claude-code">Claude Code</a>.<br> You talk; it captures, tracks, learns your taste, and reads with you.</p> <pre><code>/plugin marketplace add ZhangZhuoSJTU/YouReadIRead /plugin install you-read-i-read</code></pre> </td> </tr> </table>

<hr>

<p align="center"><i>Install once, talk to it forever.</i></p>

Highlights

<table> <tr> <td width="33%" valign="top"><b>1. Capture from any URL</b><br>arXiv, blog post, PDF, Semantic Scholar, DOI: paste a link and the agent fetches, writes a 7-section structured summary, and files it in your to-read list.</td> <td width="33%" valign="top"><b>2. Track groups and topics</b><br>Watch a research group's publications or a topic across arXiv, Semantic Scholar, and Hacker News (plus optional Twitter / LinkedIn via Apify).</td> <td width="33%" valign="top"><b>3. Discover + triage</b><br>Periodic sweeps surface new candidates. You decide per-card: add, skip, read now, or archive. Every decision is a signal.</td> </tr> <tr> <td valign="top"><b>4. Ranking that learns your taste</b><br>Accept, reject, finish, and archive shape an in-context preference model. The queue self-curates gradually, without any manual labeling.</td> <td valign="top"><b>5. Read together</b><br>Interactive reading sessions walk you through the structured summary, then drill into the raw paper text on demand with section / page references.</td> <td valign="top"><b>6. Your state, your repo</b><br>Papers, summaries, and signals live in a separate private git repo you control. The plugin tree holds none of it.</td> </tr> </table>

What you can ask it

<table> <thead> <tr> <th align="left" width="32%">You say</th> <th align="left" width="44%">It does</th> <th align="left" width="24%">Slash</th> </tr> </thead> <tbody> <tr> <td><i>"add this paper <url>"</i></td> <td>Fetches, writes a 7-section structured summary, files it in your to-read list.</td> <td><code>/add <url></code></td> </tr> <tr> <td><i>"track Stanford NLP"</i></td> <td>Watches a group's publications over time.</td> <td><code>/track-group <name></code></td> </tr> <tr> <td><i>"follow papers on LLM agents"</i></td> <td>Watches a topic across arXiv, Semantic Scholar, and HN (plus optional Twitter / LinkedIn).</td> <td><code>/track-topic <topic></code></td> </tr> <tr> <td><i>"any new papers since last week?"</i></td> <td>Sweeps every tracked source, ranks, and triages with you per-card.</td> <td><code>/update</code></td> </tr> <tr> <td><i>"let's read something"</i></td> <td>Recommends from the top of your queue and starts an interactive session.</td> <td><code>/read</code></td> </tr> <tr> <td><i>"I want to read about tool use"</i></td> <td>Searches your queue by keyword, lets you pick a match.</td> <td><code>/read tool use</code></td> </tr> <tr> <td><i>"what's on my to-read list?"</i></td> <td>Shows the queue, ranked.</td> <td><code>/todo</code></td> </tr> <tr> <td><i>"show me everything I read this month"</i></td> <td>Browses any subset of your library.</td> <td><code>/list --status read --since …</code></td> </tr> <tr> <td><i>"I finished reading X"</i></td> <td>Marks read, logs the signal.</td> <td><code>/done</code></td> </tr> <tr> <td><i>"skip this one"</i></td> <td>Archives without reading.</td> <td><code>/archive</code></td> </tr> <tr> <td><i>"push my data"</i></td> <td>Commits and pushes your private data repo.</td> <td><code>/sync</code></td> </tr> </tbody> </table>

The agent learns from every accept, reject, and read, so the queue gradually self-curates. You stay in control: it never auto-adds in default mode, and it never marks a paper read without your explicit yes.

First run

In any Claude Code session, say /sync init (or "set up You Read, I Read"). The agent will:

1. Copy defaults/config.yaml to ~/.you-read-i-read/config.yaml. 2. Ask for a data-repo remote URL. This is a separate private git repo for your personal state (papers, summaries, preferences). The plugin tree never holds any of it. 3. Optionally gh repo create --private if the remote doesn't exist yet. 4. Scaffold the data dir (with its own .gitignore so auth caches stay local).

It will not ask for an Apify token here. That's deferred until you opt into Twitter or LinkedIn.

Optional: Twitter / LinkedIn discovery (Apify)

Off by default. Twitter and LinkedIn discovery route through the Apify MCP server.

To enable:

Say "enable twitter" or "enable linkedin" in your Claude Code session.
Provide your Apify API token when asked. The first-run flow does not ask for it.
Both sources stay disabled until you opt in explicitly.

Cost (Apify ships a $5 / month free tier):

Twitter (apidojo/tweet-scraper):
$0.40 per 1 000 tweets.
Roughly $0.60 / month at default sweep rates.
Fits the free tier comfortably.
LinkedIn (curious_coder/linkedin-post-search):
$5 per 1 000 posts.
Roughly $7.50 / month at default sweep rates.
Overshoots the free tier. Mitigations: tighten tracking_sources.linkedin.max_per_sweep, sweep less often, or skip LinkedIn entirely.

Spend cap:

Default max_apify_usd_per_month: 5 per source (in ~/.you-read-i-read/config.yaml).
The plugin tracks running monthly spend in tracking/apify-spend.yaml.
Calls are skipped when the cap is approached, with a one-line notice.

Without Apify, the plugin still works:

Twitter falls back to public Nitter mirrors (free, very flaky).
LinkedIn skips silently with a plain notice.
arXiv, Semantic Scholar, and Hacker News cover the bulk of CS / ML signal regardless and require nothing.

How it's built

<table> <tr> <td><b>Skills</b></td> <td> <code>paper-curate</code> (ingest + sweep) · <code>paper-read</code> (interactive session, end-of-session prompt) · <code>paper-data</code> (list / edit / sync / first-run) </td> </tr> <tr> <td><b>Subagent</b></td> <td><code>paper-summarizer</code> for parallel summarization fanout</td> </tr> <tr> <td><b>Helpers</b></td> <td> <code>scripts/_common.py</code> (config + paper IDs + atomic file ops) · <code>scripts/manage_data.py</code> (single writer, 13 subcommands, 20 unit tests) </td> </tr> <tr> <td><b>MCP</b></td> <td><code>scripts/apify-mcp-launch.sh</code> spawns the Apify MCP server when <code>APIFY_TOKEN</code> is set; no-op otherwise</td> </tr> <tr> <td><b>State</b></td> <td>Personal data lives in a separate private git repo at <code>~/.you-read-i-read/data/</code>. Plugin tree contains zero personal data.</td> </tr> </table>

For the data layout, see docs/data-schema.md. For the agent's operating invariants, see CLAUDE.md.

License

MIT.

Summary

Install to Claude Code