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.
Sentiment Scout: momentum with receipts
Most "alpha" is a claim with nothing behind it: a post says a coin is running, a bot says sentiment is bullish, and nobody can check either. The Sentiment Scout ranks the last hour of pump.fun launches by momentum and backs every single claim with its receipt: where it came from, when it was true, and, for anything social or paid, what the chain showed at that same moment. It never trades.
Surfaces: the Scout tab on Coin Intelligence
(/coin-intel?tab=scout) · the Sentiment Scout section in every coin's
detail drawer on /coin-intel · the Scout card on trade receipts
· GET /api/pump/sentiment-scout · the sentiment_scout MCP tool · an opt-in
context block for LLM-judged sniper strategies.
Model: api/_lib/sentiment-scout.js · X client: api/_lib/x-search.js ·
Notes: api/_lib/scout-notes.js · Page module: src/coin-intel-scout.js
What a candidate looks like
{
"mint": "…pump",
"ticker": "$EXAMPLE",
"momentum_score": 71,
"score_parts": { "volume": 29.5, "buyers": 24.8, "buy_pressure": 8.4, "smart_money": 0, "curve": 9, "social": 9.7, "risk_penalty": -10 },
"evidence": [
{ "type": "volume_spike", "detail": "56.5 SOL bought in its first 90s, top 2% of 1241 launches in the window (median 0.59 SOL)",
"source": "https://three.ws/api/pump/coin-intel?mint=…", "at": "2026-09-29T06:48:22.420Z" },
{ "type": "social_mention", "platform": "x",
"detail": "19 X posts quoted this exact contract address, from 15 accounts; most-followed @… (36k followers): \"…\"",
"source": "https://x.com/…/status/…", "at": "2026-09-29T06:56:02.000Z",
"checked_against": "On-chain at 07:03 UTC: 56.5 SOL bought by 74 wallets in the first 90s; graduated to pumpswap" }
],
"caution": "1 other coin named $EXAMPLE launched in the same window; check you have the right mint.",
"market_check": { "checked_at": "…", "early_buy_volume_sol": 56.5, "early_buyers": 74, "graduated": true, "bonding_progress_pct": 100 },
"x": { "posts": 19, "authors": 15, "top_poster_share": 0.16, "most_followed": { "username": "…", "followers": 36000 } },
"posts": [ { "url": "https://x.com/…/status/…", "author": "…", "followers": 36000, "at": "…", "text": "…" } ],
"note": "Two sentences written from the evidence above, or null",
"unavailable": ["callouts"]
}
Every evidence line carries:
source: a URL that returns the fact (the Coin Intelligence row, the pump.fun coin page, the X post, or the x402 payment transaction on Solscan).at: when the fact was true. Launch-window facts are stamped at the end of the coin's observation window; the curve at the moment it was read; a post at the moment it was posted.checked_against(social, paid and news lines): the on-chain facts read in the same run, so a loud post sits right next to the chart that did or did not back it up.
Where the evidence comes from
| Type | Source | Notes |
|---|---|---|
volume_spike, fresh_buyers |
pump_coin_intel (the Coin Intelligence Engine, see trading surfaces) |
Ranked as percentiles against every launch in the same window. The median launch has one buyer, so "40x the median" would be noise. |
smart_money |
pump_coin_intel.smart_money_* |
Wallets with a profitable pump.fun record that bought early. |
graduation_approach |
live bonding curve (pump.fun) | Read at scout time. |
social_mention (X) |
X recent search for the coin's exact contract address | A ticker matches thousands of coins; a 44-character address matches one. Retweets are excluded so one account cannot count many times. |
social_mention (pump.fun) |
pump.fun callouts | pump.fun now serves these only to signed-in sessions, so they usually appear under unavailable. |
paid_signal |
sniper_coin_sentiment |
A paid Crypto Intel read over x402, linked to its payment transaction. Reads older than 6 hours are ignored. |
news_match |
pump_coin_intel.signals.news_headline |
The coin's name matches a live news story. |
A source that could not be read is named in unavailable and left out. It is
never estimated.
How it ranks and screens
- Every coin observed in the last
windowminutes (15 to 240, default 60). - A coin qualifies only with 3+ distinct buyers, top-quartile early buy volume,
and no
bundle_launchordev_dumpedflag. - The strongest shortlist is enriched (curve, X, callouts, paid reads) and
scored 0 to 100: volume 30, buyers 25, buy pressure 10, smart money 10,
curve 15, social 10, minus risk penalties. The parts are returned in
score_partsand shown behind Why 71? on the page. - One caution line per coin, deterministic and worst-first: a dumped dev, a bundled launch, a curve back near 0% after heavy early buying (the early buyers already sold), a single whale, a fresh-wallet swarm, X chatter that is mostly one account (promotion, not interest), heavy top-10 concentration, copycat tickers, a bearish paid read, sell pressure, snipers.
Social attention counts distinct accounts, not posts. Twenty posts from one account score lower than three posts from three accounts, and cost a penalty.
The Scout's read (the LLM layer)
Each board candidate can carry a two-sentence note written by the platform's
LLM failover chain (api/_lib/llm.js) from that candidate's evidence lines and
nothing else. No note is trusted as written: validateScoutNote() drops any
note that contains a number the evidence does not hold (digits or spelled out),
names a ticker other than the candidate's own or $THREE, carries a link, or
makes a promise ("guaranteed", "buy now"). A dropped note is never repaired;
the card simply shows its evidence. The page loads the evidence first and fills
the notes in when the chain answers, so a slow chain never holds the board.
Graded, not claimed
Every board candidate is recorded in sentiment_scout_reads (migration
20260929120000_sentiment_scout_reads.sql): the first sighting (score,
evidence, caution) is written once and never changes, and the latest read
overwrites the rest. Two things read it:
- The track record (
?track=1, shown above the board): of the coins the Scout flagged in the last 14 days that now have an outcome label, how many graduated or pumped, next to the same rate for every labeled launch the engine watched over the same days, and split by score band. A scout that cannot beat the base rate is noise, and this is where a reader finds out. - Trade receipts: when the Scout had flagged a coin before an agent bought it, the receipt shows the Scout's first flag with its sources, tagged "before entry". A flag recorded after the exit is dropped.
The API
# The board: the top 5 candidates from the last hour, with the track record
curl "https://three.ws/api/pump/sentiment-scout?limit=5&track=1"
# One coin, even if it would not make the board
curl "https://three.ws/api/pump/sentiment-scout?mint=<mint>"
# Evidence only, no LLM-written notes (fastest)
curl "https://three.ws/api/pump/sentiment-scout?notes=0"
| Param | Default | Meaning |
|---|---|---|
window |
60 |
Minutes of launches to rank against, 15 to 240. |
limit |
5 |
Candidates, 1 to 10. |
mint |
none | Read one coin. The response adds qualifies (would it clear the screen) and scouted (when it first made the board). |
notes |
1 |
0 skips the LLM-written note. |
track |
0 |
1 adds track_record. |
network |
mainnet |
mainnet or devnet. |
Public, CORS-open, IP rate-limited, cached for 30 seconds. Errors use the
platform shape { error, error_description }: invalid_mint, invalid_network.
The MCP tool
sentiment_scout on the main three.ws MCP server (/api/mcp), next to
trade_receipt. Arguments: mint, network, window_minutes, limit,
track_record. It returns the same board plus board_url. Read-only; like
every tool there it needs an OAuth token or an x402 payment. See
MCP.
For strategy builders: Scout context for the LLM judge
An LLM-judged sniper strategy (decision_mode: 'llm') can opt in with
llm_scout_context: true on POST /api/sniper/strategy. Its judge is then
handed the Scout's read of the coin (momentum score, every evidence line, the
caution) alongside the launch brief, built from the coin's intel row, a cached
one-hour window, the latest paid read and any X posts the board already
recorded, with no network call on the hot path. The read only exists once a
coin has been observed, so in practice it feeds intel_confirmed arms. It is
off by default so every running experiment keeps asking the question it was
asking; an arm with it on is marked + Scout on
/sniper/experiments.
X search budget
X recent search is metered per app per 15 minutes. The Scout only searches
shortlisted candidates, caches each coin's posts for 10 minutes (a failed read
for 1 minute), and stops searching once the remaining quota drops below a
reserve, so it can never exhaust the quota other X features share.
Credentials: X_BEARER_TOKEN, or the app key pair X_API_KEY / X_API_SECRET
from which an app token is minted. Without them, and without the fallback
below, X posts are listed as unavailable and everything else works.
Failover: X search through xAI
When our own X bearer cannot answer (no credentials, a rejected token, a 429,
the reserve reached, X down), api/_lib/x-search.js makes one call to xAI's
Responses API (POST https://api.x.ai/v1/responses) with the built-in
x_search tool and a strict JSON schema, then parses the answer into the same
post shape. The candidate's x.rung field says which rung served (bearer or
xai).
| Env var | Meaning |
|---|---|
XAI_API_KEY (or GROK_API_KEY) |
Enables the rung. Unset means the rung never runs. |
XAI_X_SEARCH_DAILY_CAP |
Calls per UTC day, default 200. 0 turns the rung off. Reaching it fails with reason xai_daily_cap. |
XAI_X_SEARCH_MODEL |
Model id, default the flagship Grok in api/_lib/chat-models.js. |
xAI bills x_search per call, so the rung only runs after the bearer path
fails. Post text is untrusted: a post is kept only if the tool itself cited its
status URL and its text contains the mint, so an invented post is dropped.
Each served search logs one x_search JSON line (rung, latency, post count)
and per-day counts per rung are readable with xSearchStats().
Second rung: xAI's x_search tool. When the bearer rung above cannot
answer (not configured, rate limited, or X itself down) and an xAI key
(GROK_API_KEY or XAI_API_KEY) is configured, searchMintPosts falls over
to a second, independent read of the same data: xAI's x_search tool on its
Responses API (POST https://api.x.ai/v1/responses). Grok runs the search
server-side and is instructed to answer with a strict JSON array of matching
posts, parsed defensively into the same post shape the bearer rung returns
(engagement counts and author detail the tool does not expose are left null
rather than invented). That answer is untrusted model output, never
instructions. Every xAI call is billed per call, so a daily cap
(XAI_X_SEARCH_DAILY_CAP, default 200, counted in Postgres app_settings so
it holds across instances) stops the rung for the rest of the day once
reached; callers then see the same unavailable shape as any other outage.
Code: searchMintPostsViaXai / parseXaiSearchPayload in
api/_lib/x-search.js.
Related
- Trading surfaces: Coin Intelligence and the launch data every volume and buyer line is measured from.
- Trade receipts: where a Scout flag shows up on a trade.
- Copy Coach: the guided first copy, which links here.