org.sentimentfxPublished by sentimentfx.orgOfficial publisherv1.0.0Updated Oct 24, 2026
FinBERT news sentiment and price history for 42 crypto, FX, stock, ETF and commodity tickers.
SentimentFX exposes FinBERT-scored news sentiment, daily prices and sentiment-price correlation for 42 crypto, FX, stock, ETF and commodity tickers over a…
Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.
Paste this prompt into Claude Code, Codex, Cursor or another agent — or into SourceWeft chat. The agent reads this server's install guide, shows you its publisher, endpoint and sign-in, and adds it to your client once you agree. Keys never go into a command.
Read https://sourceweft.com/mcp/org.sentimentfx%2Fmarket-sentiment/install.md and connect the MCP server it describes to my client. Before changing any config, show me its publisher, endpoint and sign-in method and wait for my OK. Never put an API key in a command or in a file I would commit; ask me where it should come from.Add this to your client's mcpServers config.
{
"mcpServers": {
"sentimentfx": {
"type": "http",
"url": "https://api.sentimentfx.org/mcp",
"headers": {
"X-API-Key": "<YOUR_API_KEY>"
}
}
}
}A news-sentiment dataset for 42 markets across crypto, FX, stocks, ETFs and commodities. SentimentFX scrapes financial headlines, scores each one with FinBERT (a finance-tuned transformer model), and stores the scores next to daily and intraday prices. You can read the data on a dashboard, through a metered REST API, or as tools over MCP for Claude and other AI clients.
The product is the scored archive and live feed. It is not a price forecast: the dashboard shows correlation and backtest results so you can judge the signal yourself, and it makes no claim that sentiment predicts returns.
Live: sentimentfx.org · app.sentimentfx.org · developers.sentimentfx.org · status.sentimentfx.org
positive_prob - negative_prob, giving -1 to +1, shown as 0–100 in the UIPrice currency depends on category. Crypto is quoted in GBP (BTC-GBP on Yahoo). FX pairs are raw exchange rates. Stocks, ETFs and commodity futures are in native USD, and no GBP conversion is applied.
Alerts only fire for tickers whose backtest passes a quality gate: positive out-of-sample net return, and at least 70% of walk-forward folds positive. The gate is refreshed daily.
The bulk archive (every scored headline and price row, as CSV or JSONL) is sold separately. Ask through the dataset form on the developer portal.
api.sentimentfx.org/v1/)Authenticate with an X-API-Key header. Full docs are at developers.sentimentfx.org.
Social posts (Reddit, StockTwits, X) are collected but kept separate. Every sentiment number is news only by default, and sources=social returns the social posts on their own. Calls are counted after the query, on rows actually returned. Nothing is billed per call: a key that uses up its allowance gets a 402 insufficient_quota until it refills, and a higher allowance can be requested. A 404 costs nothing, and out-of-range parameters are rejected with a 422 before any work is done. Every response carries X-Quota-* and X-RateLimit-* headers. Errors use one envelope: {"error": {"type": "...", "message": "..."}}.
The Sentiment Index is public and needs no key: GET https://api.sentimentfx.org/sentiment-index. Like every other sentiment number, it's built from news only.
api.sentimentfx.org/mcp)The same data is exposed as six Model Context Protocol tools: list_tickers and get_usage (both free), plus get_sentiment, get_summary, get_prices and get_correlation, which bill exactly like their /v1 equivalents. It uses streamable HTTP with the same X-API-Key header, or Authorization: Bearer <key> for clients that can't set custom headers (the Claude API's MCP connector is one). The registry manifest is server.json, and publishing steps are in docs/publishing-mcp-registry.md.
Scheduled jobs (APScheduler, inside the API process):
FinBERT only runs on headlines it hasn't seen. The model (~400 MB) downloads and loads on the first scoring call, not at import.
Dormant sources. GNews, StockTwits and X/Twitter fetchers are still in the code but switched off. Set GNEWS_ENABLED, STOCKTWITS_ENABLED or X_ENABLED to true to turn one back on. StockTwits was turned off because Cloudflare blocks the server's IP and retail chat scores poorly with a news-tuned model.
There is no .env.example. Create backend/.env with the variables below.
Set VITE_SUPABASE_ANON_KEY in frontend/.env to enable login (VITE_GA_ID is optional). The API base URL is hard-coded to https://api.sentimentfx.org in src/lib/constants.js, so a local frontend talks to production unless you change it.
Use python -m pytest, not bare pytest: tests/ has no __init__.py, so bare pytest can't import app. The tests are smoke-level only. They check that the app boots and the core routes answer, and they need a reachable DATABASE_URL. CI runs them against a throwaway Postgres on every push to main and every PR. There is no frontend test suite.
The range is -1 (most negative) to +1 (most positive), and the UI shows it as 0–100. Neutral headlines land near 0 instead of being forced to exactly 0, which was the problem with the old label × confidence method. FinBERT scores the title only, even when the full article body is stored.
Procfile + nixpacks.toml at api.sentimentfx.org. Starts with uvicorn app.main:app --host 0.0.0.0 --port $PORT. Runs on CPU only.frontend/, landing/, developers/), each auto-deploying from main.Source: README.md at commit 0357f97
List every asset SentimentFX tracks. Cheap (free — doesn't hit the billing meter). Returns two groups: `primary` (5 crypto + 7 FX pairs — the ones with full sentiment coverage across the primary RSS feeds), and `background` (US equities, ETFs, commodity futures — coverage is thinner but real). Use the exact ticker string from either group as the `ticker` arg to the other tools.
Takes no parameters.
Introspect your API key: included allowance, remaining credits, and when they refill. Free keys get 1,000 calls per 30-day window starting at their first billable call (`resets_at` is null when no window is running); paid plans refill on the 1st of each month. Free — doesn't hit the billing meter. Useful before a large batch of `get_sentiment`/`get_summary` calls to check you have credit headroom.
Takes no parameters.
Return the most recent FinBERT-scored news headlines for `ticker`. Each headline has a `sentiment_score` in [-1, +1] (positive_prob - negative_prob), a categorical `sentiment_label`, and a `source_kind` (news / reddit / stocktwits / x / youtube). Costs 1 API credit per 25 headlines actually returned — same billing as GET /v1/sentiment/{ticker}. An empty result costs nothing. `limit` is capped at 100. Social posts are kept separate from news: the default `sources="news"` returns editorial news, `"social"` returns only Reddit, StockTwits and X posts, and `"all"` returns both, labelled by `source_kind`. Titles come back reverse-chronological so the first item is the freshest.
limittickerrequiredsourcesReturn daily aggregated news sentiment for `ticker` over the last `days`. Each entry has `avg_sentiment` (unweighted mean of that day's scores), `article_count`, and a directional `label` (positive / negative / neutral, thresholded at ±0.1). `sources="social"` returns a separate daily average of Reddit, StockTwits and X posts; news and social are never averaged together. Costs 1 API credit per day actually returned — same as GET /v1/summary/{ticker}. A window with no coverage costs nothing.
daystickerrequiredsourcesReturn daily close prices for `ticker` over the last `days`, in the ticker's native currency -- see the `currency` field on the response. Costs 1 API credit per day actually returned — same as GET /v1/prices/{ticker}. Prices come back reverse-chronological. Crypto is GBP (yfinance BTC-GBP etc.), FX pairs are a raw exchange rate in the pair's native convention (e.g. USDJPY is yen per dollar), and everything else (stocks/ETFs/commodity futures) is native USD -- there is no currency conversion.
daystickerrequired180-day Pearson correlation between daily sentiment shifts and next-day price returns for `ticker`. Returns Pearson `r`, `p_value`, `n_days` overlapping, a 95% confidence interval (Fisher z), and a categorical `strength` (strong / weak / inconclusive). Costs 1 API credit — same as GET /v1/correlation/{ticker}. Requires ≥30 overlapping day-pairs. Under that, returns a `note` field explaining what's missing so a caller can suggest waiting or switching to a higher-coverage ticker.
tickerrequired90% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
Declare a license in the repository or the registry entry.
# Get a free key (1,000 calls every 30 days, no card)curl -X POST https://api.sentimentfx.org/api/keys/generate \ -H "Content-Type: application/json" \ -d '{"email": "[email protected]"}'
# Use itcurl https://api.sentimentfx.org/v1/summary/BTC?days=7 \ -H "X-API-Key: sfx_your_key_here"{ "mcpServers": { "sentimentfx": { "type": "http", "url": "https://api.sentimentfx.org/mcp", "headers": { "X-API-Key": "sfx_your_key_here" } } }}RSS feeds (15 min) ─┐ScraperAI replay ───┼→ scraper.py / ai_sources.py → dedup by URL → FinBERT → headlinesHN Algolia (admin) ─┘
yfinance (hourly) → prices.py / candles.py → prices, candles
headlines + prices → /dashboard, /candles, /correlation, /backtest … → React dashboard → /v1/* and /mcp → API and MCP clients → /sentiment-index → public index pagecd backendpython -m venv venvsource venv/Scripts/activate # Windows (Git Bash); venv/bin/activate on macOS/Linuxpip install -r requirements.txtpython -m uvicorn app.main:app --reload # http://localhost:8000cd frontendnpm installnpm run dev # http://localhost:5173cd backendpip install -r requirements-dev.txtpython -m pytest tests/ -vscore = positive_probability - negative_probability