Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
The X post archive
Every post our own X accounts publish, kept and measured, so "what actually lands with our audience" is a query instead of an opinion.
This is internal marketing infrastructure. It is not a product surface, it does not touch a user's account, and nothing here posts anything: it only reads what was already published.
Why it exists separately from x_posts
The platform already has an x_posts table, but it holds what a three.ws
user published through the platform (POST /api/x/post), keyed to their
account. The @trythreews timeline is a different corpus: it predates that
feature, most of it was written by hand, and it is the only body of evidence we
have about what our audience responds to. Mixing the two would mean answering a
marketing question by scanning somebody else's agent's posts.
What is stored
Three tables, created by api/_lib/migrations/20260814120000_x_account_archive.sql:
| Table | One row per | Why |
|---|---|---|
x_account_imports |
snapshot file ingested | Hashed, so the same file cannot be counted twice. |
x_account_posts |
post | The post and its latest observed metrics. |
x_account_post_snapshots |
post, per import | Engagement is a time series. "40 likes on day one, 187 by day three" is two measurements, not a correction. |
The raw snapshot files in data/x-archive/ are the durable copy. The database is a queryable index over them, and can be rebuilt from them at any time.
The commands
npm run x:archive:import # snapshot files -> Postgres, idempotent
npm run x:archive:refresh # exact metrics from the X API -> a new snapshot file
npm run x:archive:analyze # rebuild the engagement report
Typical loop after a fresh scrape lands in data/x-archive/:
npm run db:status # confirm the archive migration is applied
npm run x:archive:import
npm run x:archive:analyze
analyze reads the database when DATABASE_URL is set and falls back to the
archive files when it is not, so the report can always be regenerated from a
clean checkout.
The report
docs/x-archive/trythreews-engagement.md is generated, never hand-edited. It carries:
- headline totals and the engagement distribution (median, p75/p90/p99, and the share of all engagement earned by the top decile),
- the top posts by raw engagement, by engagement rate (which normalizes for reach, so it ranks writing rather than luck), and by replies,
- lift tables: for each measurable signal (image, video, link card, mention, question, length band, subject), the group's median engagement divided by the corpus median, shown against the posts without that signal,
- timing by hour and weekday, cadence by month, and who we amplify.
Lift is computed against the median, not the mean, on purpose: with the top 10% of posts earning roughly half of all engagement, a mean is a report about one viral post.
X Articles and what readers said back
Articles are the long form, and the useful signal is in the replies, not just the
counts. npm run x:articles:pull walks the timeline through the X API, keeps every
post that carries an Article (title, full body, exact metrics), and collects the
readers' replies and quotes for each one. The file format and flags are in
data/x-archive/README.md.
docs/x-archive/trythreews-articles-readers.md is the hand-written reading of that pull: which Articles reached people, what readers asked for in their own words, and what that means for the next Article. Unlike the engagement report it is not generated, so update it after a fresh pull.
Two measurement rules worth knowing
Reposts are not our posts. A profile scrape returns the whole timeline, and
a repost or quote keeps the original author's permalink and engagement. The
scraper's own isRetweet flag was false on all 145 reposts in the first
archive, so authorship is read from the permalink instead. Only self-authored
posts are measured; who we amplify is reported separately.
A scraped zero is not always a zero. X renders like counts lazily, so a
scroll-based scrape captures real posts at "0 likes": 56 of the first 214 own
posts, including one with 172K views and 92 replies. Those are archived, named
in the report, and excluded from every median rather than dragging it down.
npm run x:archive:refresh replaces them with exact public_metrics from the
X API (X_API_BEARER), after which nothing is excluded.
View counts past 999 are abbreviated by X itself (6.3K), so scraped view
numbers carry two significant figures. Both the label and the parsed integer are
stored, with views_exact recording which is which; the API refresh returns
exact impressions.
Adding another account
Nothing is hardcoded to one handle. Drop <handle>-<date>.json into
data/x-archive/ in the same shape (profile, scrapedAt, tweets[]) and run
the three commands; the report is written per handle.
Code: scripts/x-archive-lib.mjs (normalization + analysis, covered by tests/x-archive.test.js), scripts/x-archive-import.mjs, scripts/x-archive-refresh.mjs, scripts/x-archive-analyze.mjs.