New papers, code, models and data that could help read the scrolls, collected from public sources around the clock and scored for relevance. Open to everyone, with an API and an MCP server.
Items per hour, stacked by source.
Items touching each topic. Click a bar to filter by it.
How scores spread. Darker is a higher band; the dashed line is your minimum.
Items per source in this window. Click a row to filter.
Which sources feed which topics. Darker cells hold more items.
Ranked by distinct pieces of work, each worth its best score: many small related items count as one.
The best three items per topic in the window and view you picked above, score 35 or more, each item listed once. Scores are a triage aid, not a judgment: open the item before you act on it.
| Source | Every | Last check | Next | State | Read | Kept | New | Stored | Time |
|---|
| Finished | Source | State | Read | Kept | New | Time | Notes |
|---|
Everything this page shows is open JSON, CSV and RSS over HTTPS, and the whole radar is also an MCP server, so a script, a spreadsheet, a feed reader or an agent can use it. No key, no login and no GPU quota: anyone can call it, logged in or not.
https://sharpmeow-vesuvius-radar.hf.space/mcpMCP endpoint (Streamable HTTP, POST): https://sharpmeow-vesuvius-radar.hf.space/mcphttps://sharpmeow-vesuvius-radar.hf.space/api/items?since=1dItems from the last day (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/items?since=1d
Papers and methods from the last week, score 35 or more (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/items?since=7d&view=research&min_score=35
Search by meaning (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/items?since=30d&q=ink+detection+in+x-ray+ct&by_meaning=true
The strongest items per topic (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/briefing?since=7d
Counts behind the charts (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/overview?since=1d
The collector: each source's last and next check (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/sources
The last week as CSV: https://sharpmeow-vesuvius-radar.hf.space/api/export.csv?since=7d
RSS feed: last week, score 60 or more: https://sharpmeow-vesuvius-radar.hf.space/feed.xml
Reader views and their rules (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/views
Sources, keywords, weights and anchor sentences (JSON): https://sharpmeow-vesuvius-radar.hf.space/api/config
OpenAPI page with every parameter: https://sharpmeow-vesuvius-radar.hf.space/docsThese follow the view, lookback and filters you picked on the Radar tab.
Every parameter, with a form to try it, is on the OpenAPI page.
The radar is an MCP server at /mcp (Streamable HTTP). No key or header is needed.
Claude Code:
claude mcp add --transport http vesuvius-radar \
https://sharpmeow-vesuvius-radar.hf.space/mcpAny client that takes a JSON config:
{"mcpServers": {"vesuvius-radar": {
"type": "http",
"url": "https://sharpmeow-vesuvius-radar.hf.space/mcp"}}}| Tool | What it does |
|---|---|
search_items | Filter a window by words or rank it by meaning |
briefing | The best few items per topic |
get_item | One item, its score breakdown and its nearest neighbors |
find_similar | Items closest in meaning to one item |
overview | Counts by source, band, topic, kind and person |
source_status | Each source's last and next check |
check_now | Ask every source for its newest items |
list_views, get_settings | Reader views; keywords, anchors and schedules |
Standard library only, no key.
import json, urllib.parse, urllib.request
BASE = "https://sharpmeow-vesuvius-radar.hf.space"
def get(path, **params):
url = BASE + path + "?" + urllib.parse.urlencode(params)
return json.load(urllib.request.urlopen(url))
week = get("/api/items", since="7d", view="builders", min_score=35, limit=100)
for i in week["items"]:
print(i["score"], i["source"], i["title"], i["url"])
# rank the last month by meaning
hits = get("/api/items", since="30d", q="tracing papyrus fibers", by_meaning="true")curl -s "https://sharpmeow-vesuvius-radar.hf.space/api/items?since=6h&min_score=35" \
| python3 -m json.toolIn Google Sheets or Excel, import the CSV link above (for example =IMPORTDATA("...export.csv?since=7d")). A feed reader takes the RSS link.
Only public pages and APIs, logged out, on a schedule per source (the Sources tab shows when each last ran). It keeps 30 days, so changing the lookback is instant.
Anyone can press Check now; to stay polite to these services there is one check at a time for the whole Space and a few minutes between checks. Chat servers and social media are not read. When the Space has been asleep it catches up on the time it missed at the next check.
Every item gets a keyword score and a meaning score from 0 to 100; a small learned model reads them with the text to set the final score. The page shows all three.
Keywords. Scroll terms (Vesuvius, Herculaneum, PHerc, papyrus, Volume Cartographer and similar) add the most. In each topic a strong term counts on its own and a weak term counts half, and only when a context term such as 3D, CT or volume is present too. Terms like LLM or chatbot take points off. Each item lists every term group and its points.
Meaning. A small sentence-embedding model (BAAI/bge-small-en-v1.5, run on the Space's CPU) compares the item's title and the start of its summary (600 characters) with a set of sentences that describe the work the scrolls need, and reports the closest one. That catches items that never use our words. After the Space restarts, its first check reads every listing by meaning again, which takes several minutes on a small CPU; the Sources tab shows its progress, and items already stored stay on the page meanwhile.
Final score. A small learned model sets it: a ridge regression on the item's embedding plus its keyword and meaning scores, source, kind, stars and likes, fitted to blind ratings (0 to 100) of what each of several hundred collected items is worth to scroll work, and calibrated so each band holds about as many items as the ratings put there. On held-out items it ranks them much closer to the ratings than keywords and meaning alone did (figures in the Learned model card below). Items with no text of their own get the same model without the embedding. Which items are kept is still decided by keywords and meaning: the higher of the keyword score and a blend of both, with a long text the model reads as unrelated held below high unless it comes from ScrollPrize/villa. The same embedding powers search by meaning and "More like this".