Overview
Augur (pronounced AW-ger—from ancient Rome, meaning to predict or indicate what may happen) is Sentimentor's integrated AI research-agent system—not merely a branded chat interface. Its agent loop plans an investigation, dynamically selects from 41 first-party tools, runs independent calls in parallel when useful, evaluates the evidence over multiple rounds and synthesizes a grounded answer. Ask it for a quote, a screener, a sentiment read, a multi-ticker signal dashboard, a defined-risk options play, a backtested strategy, a full workflow, or a saved research report; Augur picks the right tools, runs them against current market and public-record data, and renders the results as typed cards inline with its response.
- Grounded, not hallucinated. Every numeric claim comes from a tool call you can inspect — the chip strip under each turn shows what ran and whether it succeeded.
- Multi-round reasoning. Complex asks ("screen for momentum then build me a strategy for the top pick") execute across several rounds; the agent re-plans after each batch of tool results.
- Typed result cards. Snapshots, sentiment gauges, news feeds, screener tables, radar / scatter / Greeks-alignment dashboards, defined-risk options plays (with P&L curves + heatmaps + Greeks), generated PineNode strategies, generated workflows — each renders as its own UI component, not a wall of JSON.
- Reads URLs. Paste a Hunterbrook / Muddy Waters investigation URL, an SEC filing, an 8-K, or any public article and Augur will fetch and read it before answering (see
fetch_url). - One click to act. Generated strategies open in the PineNode editor; generated workflows save to your library; saved reports land in
/reports.
Quickstart
Open the chat panel from the right-rail mascot. A useful prompt names the objective, symbol or universe, timeframe, evidence to compare, and the output you want.
promptLook up NVDA.
promptAnalyze NVDA over the last 90 days. Compare price and volume with current options flow, IV rank, sentiment, institutional ownership, and recent news. Identify a testable momentum hypothesis, mark missing evidence, and generate editable PineNode rules.
Other useful starting points:
promptWhat's the sentiment on NVDA right now?
promptFind me 5 momentum names with RSI 50–75 and volume above 5M, then generate a PineNode strategy for the highest-scoring one.
promptCompare NVDA, AMD, INTC, GOOGL and META across options activity, sentiment, IV rank, dark-pool conviction, and momentum. Show me a radar and mark missing evidence.
promptBuild me a bull call spread on NVDA for the Aug 21 expiry, 145/155 strikes, quantity 5. Show P&L curve, heatmap, and Greeks.
promptBuild me a daily workflow that screens for breakouts at 9:35 AM and emits a strategy per pick.
For each turn you'll see a strip of small chips listing the tools Augur invoked, color-coded by status — blue while running, green on success, red on error. Tools that ran multiple times in a single turn collapse into one chip with a ·Ncount suffix so the strip stays scannable even on multi-round reasoning.
Anatomy of a turn
A "turn" is one user prompt and the assistant response that follows. Inside a single turn, the agent may invoke many tools — often in batches that go out in parallel — and run the loop several times before it composes a final answer.
Reasoning rounds
Each pass through "think → call tools → read results → think again" is a round. Simple questions resolve in 1–2 rounds; long-form research can stretch to 5–8. The streaming bubble shows the current round counter (Streaming · round 4) so you have an honest signal that the agent is still working.
Streaming + tool chips
Text deltas stream as the model produces them. Tool calls show up as chips beneath the streaming bubble with three states:
- running — call dispatched, awaiting result.
- ok — result returned, rendered inline.
- error — upstream API failed. Hover for the message.
Duplicate tool names collapse: if the agent callsget_stock_snapshot three times across two rounds, you see one stock snapshot · 3 chip rather than three separate badges.
Tools catalogue
Augur ships with 41 first-party tools, grouped by purpose. The agent reads the catalogue every turn and picks tools dynamically — you don't need to invoke them by name. The list below is for understanding what coverage exists.
Quotes + screeners
get_stock_snapshot— live OHLC, % change, volume.search_symbols— resolve a name to a ticker.get_top_movers— gainers / losers / most-active.get_market_overview— market-wide health summary.get_market_movers_extended— premarket + after-hours movers.get_screener_results— quant screen on equities.screener_advanced— multi-factor screen with custom thresholds.get_historical_bars— OHLC history at any timeframe.get_cross_asset_market_recap— dated equity, crypto, forex and index recap.
Sentiment + prediction
get_sentiment_score— Sentimentor composite (0–10) per ticker.get_sentimentor_prediction— proprietary directional read with confidence.
Order flow
get_options_flow— unusual contract activity, option marks, Greeks and call/put composition. Snapshot-only: no verified sweeps or buyer/seller direction.get_dark_pool_levels— off-exchange institutional prints.get_volume_delta— buy vs sell pressure imbalance.
Historical options metrics
Historical options context that live flow tools alone can't answer — is today's IV expensive vs the ticker's own history? Is the put/call ratio bending bullish or bearish over the last N days? Both backed by the daily options-metrics cron (top-500 by dollar volume + user-watchlist tickers).
get_iv_rank— 0–100 percentile of the current ATM 30-DTE IV against the ticker's own history (up to 252 days). Returnslow_confidence: truewhen the sample window is short (30–89 days); ramps to true 52-week rank at 252+ days. Untracked tickers returniv_rank: nullwith a note suggesting a watchlist add.get_pc_ratio_trend— put/call ratio trend over the last N days (default 5). Returnsdirection: 'rising' | 'declining' | 'flat', per-day history, and slope. Semantics: declining is bullish (rotation into calls); rising is defensive.
Fundamentals
get_financial_metrics— P/E, EPS, margins, dividends.get_analyst_recommendations— consensus + target prices.get_institutional_ownership— 13F holders + concentration.get_related_companies— peer-set discovery.get_comparable_companies— valuation comps.
Calendar
get_economic_calendar— CPI / FOMC / NFP / GDP releases.get_earnings_calendar— upcoming earnings dates.get_ipo_calendar— pricings + first-trade dates.get_ex_dividends— declared dividends.get_corporate_activity— splits, M&A, buybacks.
CFTC + insider
get_cftc_commitments— commercial vs speculator positioning.get_insider_transactions— recent Form 4 filings.run_form4_conviction_scan— evidence-backed Form 4 conviction screening with executive compensation, ownership materiality, Altman Z-score, current ratio, debt-maturity and credit-facility checks.get_insider_sentiment— net buy/sell pressure.get_congressional_trades— normalized House, Senate and presidential disclosures.Public-record research— lobbying, government contracts, 13F filings, patents and related disclosures.
Web fetch
When you paste a URL (a short-report investigation, a press release, an 8-K filing, a blog post), Augur fetches and reads it before answering. This closes the "I can't visit URLs" gap that Claude's base model has by default.
fetch_url— retrieves a public web page and returns its main text content. Public URLs only; SSRF-guarded against private / loopback / cloud-metadata hosts (localhost, 10.x, 169.254.169.254, etc.). Text-y content types only (HTML, JSON, XML, plain text); binaries (PDF, images) rejected. Response capped at ~24KB of extracted text with a 15s timeout. Follows up to 5 redirects.
Typical flow: user pastes a Hunterbrook short-report URL → Augur calls fetch_url → extracts the target ticker + thesis → runs get_market_news, get_options_flow,get_insider_transactions on the target in parallel → delivers a real analysis of the short thesis with corroborating or refuting data.
Generation tools
These are the tools that produce artefacts — not data, but objects you can save and act on.
generate_pinenode_strategy— outputs PineNode source code with parameter bindings.generate_workflow_definition— outputs a workflow graph (nodes + edges) ready to save.generate_dashboard— emits a workflow visualization payload: a radar chart of signal strength (0–100) across common dimensions for each pick, an optional catalyst-timing scatter, and a Greeks-alignment panel only when signed, execution-level positioning data exists. Contract snapshots do not qualify. Renders as the top workflow-report card.generate_options_strategy— emits a structured priced, same-expiry, fully hedged options play for one ticker (vertical spread, iron condor, or iron butterfly). Rendered per-pick as anOptionsStrategyCard with metrics tiles (Net Debit, Max Profit, Max Loss, estimated Probability of Profit, Breakeven), a P&L payoff curve at expiry, a P&L heatmap across price × time, and Greeks tiles (client-side Black-Scholes). Debit, max profit/loss, breakevens and expiry payoff are recomputed server-side from contract marks. Naked shorts and unpriced legs are rejected.save_analysis_report— persists the conversation's findings as a citable report.
Result cards
Augur doesn't dump tool JSON into the chat. Each result tool has a typed renderer that turns the response into a glance-friendly card. The cards live in the message body and survive scrollback.
Stock snapshot
Big price, signed % change, signed $ change, OHLV in a four-up grid. Two anchor buttons at the bottom open/chart/{sym} and/stocks/{sym} in new tabs so you can deep-dive without losing the conversation.
Sentiment gauge
Score 0–10 with a horizontal gauge bar, category label (bullish / bearish / neutral), and a single-line rationale pulled from the tool response.
Screener table
Compact table of rows with ticker, name, key metrics. Click a row to open the per-symbol stock page in a new tab.
News
Each article surfaces title, source, time, and a per-article sentiment score (1–10). Hits the article URL in a new tab on click.
Signal dashboard
When the agent calls generate_dashboard, a full signal-dashboard card renders inline in the chat bubble — the same card that workflow run reports show at the top. Three panels:
- Radar — signal strength per dimension for each pick (0–100). One polygon per ticker, layered. Useful for multi-ticker signal comparisons across a shared set of dimensions (e.g. options activity / sentiment / IV rank / momentum / dark pool).
- Scatter (optional) — classifies picks along two continuous axes. Typical: X = catalyst timing (days), Y = signal strength (0–100), one dot per pick colored by category (Earnings / M&A / Product / Macro / etc.).
- Greeks alignment (conditional) — shown only when a time-aligned execution source establishes buyer/seller direction. A contract snapshot cannot support signed-flow or alignment verdicts.
Example prompt: "Compare NVDA, AMD, INTC, GOOGL and META across options activity, sentiment, IV rank, dark-pool conviction, and 30-day momentum. Show me a radar."
Options play
When the agent calls generate_options_strategy, a full defined-risk options card renders inline — the same decision-oriented view that workflow reports mount per pick:
- Header + metric tiles — strategy name, direction, DTE, current spot, then eight tiles: Net Debit (or Net Credit), Max Profit, Max Loss, estimated Probability of Profit, Breakeven, Expiry, estimated Expected P&L, Avg IV.
- P&L curve tab — Recharts area chart of the payoff at expiry across a price range slider (±2% to ±50%). Green shading above zero, red below; dashed reference line at current spot; yellow dots at breakevens.
- P&L heatmap tab — price × time grid (15 prices × 10 dates) with cells colored by live P&L (Black-Scholes on the client). Lets you see how the play behaves at any (S, T) combination, not just at expiry.
- Greeks tab — net position Delta / Gamma / Theta-per-day / Vega tiles + per-leg Greeks table, all computed live via Black-Scholes.
- Legs tab — the actual contract legs, one row each (action / type / strike / expiry / quantity / premium / IV).
Workflow plays are built from priced, same-expiry contracts. If a valid pair is unavailable, the report explains why instead of inventing a play. Naked shorts and unpriced legs are rejected.
Example prompt: "Build me a bull call spread on NVDA for the Aug 21 expiry, 145/155 strikes, quantity 5. Show P&L curve, heatmap, and Greeks."
PineNode strategy
When the agent calls generate_pinenode_strategy, a card appears with the generated strategy name, a short description, and an Open in editor button that drops you into /pinenode with the strategy pre-loaded. Useful for one-shot "build me an RSI bounce strategy on AAPL" flows.
Workflow definition
generate_workflow_definition renders a save button: review the proposed node graph, hit save, the workflow lands in your library at /workflows ready to run. See the Workflows docs for what each node type does.
Generating PineNode strategies
Strategy generation is the most common artefact flow. A well-formed ask names the symbol, the rule, the timeframe, and any guards.
promptGenerate a PineNode strategy for AAPL on the 5-minute that buys an RSI(14) crossOver(30), confirms with a +0.5σ Bollinger Band move, and sells on crossUnder(70) — only during the regular session.
Augur calls generate_pinenode_strategy with the decoded rule set and renders a strategy card. Open the editor to backtest, debug, or tweak parameters — the generated code is real PineNode, not a stub.
Generating workflows
A workflow is Augur's answer to open-ended quantitative analysis. Instead of hiding the investigation inside one large generated program, Augur decomposes it into an ordered graph of data calls, custom steps, transforms, scoring, ranking and conditional operations. The run page preserves a Code view and result for each tool-backed operation, then the required terminal step emits complete PineNode strategy source for every surviving candidate.
Workflows are repeatable research playbooks. Ask Augur to build one and it composes a node graph from the catalogue described in the workflows docs.
promptBuild a daily workflow that: (1) runs the screener for momentum tickers with RSI 50–75 and volume >5M, (2) pulls a sentiment score and dark-pool levels for each, (3) ranks by composite score, and (4) emits a PineNode strategy for the top 3.
The saved workflow appears in/workflows, ready to run on demand or schedule (Elite tier). The required terminal node is alwaysoutput.emit_strategy — strategies are the workflow's artefact contract.
Build-as-workflow mode
The Build as workflow chip in the composer tells Augur to save its work as a runnable, saveable workflow instead of answering as a one-shot conversation. Three states:
- Off (default) — Augur answers now, no workflow is created. Best for exploratory questions.
- ON (auto) — as you type, the composer watches for workflow-intent phrasing ("recurring", "scanner", "every morning", "build me a workflow", etc.). When it fires, the chip lights up as
· AUTO. Same heuristic the backend uses, so the chip promises what the agent will actually do. - ON (manual) — click the chip before sending to force workflow-build regardless of prompt wording. The chip shows
· ON. This is the escape hatch for prompts the auto-detect misses.
Forcing a workflow on a prompt that doesn't auto-detect
Type your request as you normally would, then click Build as workflow before hitting send. The chip locks ON and Augur persists the result as a runnable, re-usable workflow even though the wording didn't include the usual keywords.
promptFind mid-cap semiconductors with insider buying in the last 30 days and a positive sentiment shift.
This reads as one-shot to the auto-detector (no "every", no "recurring", no "scanner"). If you want it saved as a workflow you can re-run weekly, click Build as workflow first.
Conversely, clicking the chip while it's auto-detected turns it OFF and locks the state until the next send — useful when the heuristic tripped but you actually just want a one-time answer. Auto-detection resets whenever the composer clears (fresh draft = fresh chance to auto-detect).
Research Evals: measure predictions forward
Backtests answer how a fixed rule would have behaved on a historical series. They do not, by themselves, prove that an AI ranking was knowable at the time or will outperform later. Research Evals provide a separate, immutable ledger for that question. Open /research-evals, paste Augur's ranked output, and seal it before outcomes are known.
promptRank 10 S&P 500 stocks by their probability of outperforming SPY over the next 63 trading days. Give each a probability, concise thesis, and the evidence used so I can seal the result as a Research Eval.
A v1 seal captures:
- the exact prompt, universe, model/provider/version and prompt version;
- the as-of time, evidence cutoff and point-in-time evidence manifest;
- ranked symbols, forecast probabilities, benchmark, horizon and transaction-cost assumption;
- server-captured entry closes and a SHA-256 checksum of the record.
A scheduled scorer appends provisional returns as bars arrive. It marks results final only after the full trading-day horizon, then reports candidate return, benchmark return, alpha, hit rate and Brier score for supplied outperformance probabilities. The original prompt, picks and entry snapshots have no edit or delete API.
Five supported tests
- Blind forward selection: seal an unchanged ranking request now and measure it later.
- Historical replay: declare a past cutoff and preserve every source used in the replay.
- Regime test: tag equivalent runs across bull, bear, volatile, rate and sideways conditions.
- Human/tool comparison: freeze the same question and score return, time, reproducibility and false positives consistently.
- AI ablation: preserve a simple rules baseline to test whether AI adds measurable value beyond the interface.
Saved reports
Reports persist a conversation's findings into a structured note you can reference, share, or re-open. Ask Augur to "save this as a report" once you've covered the analysis you want preserved, and it calls save_analysis_reportwith the synthesised content. The saved report card links to the persistent URL.
Conversation context
When you open Augur on a ticker-specific page (/stocks/AAPL, /chart/NVDA), the symbol gets passed in as turn-level context. You can ask follow-ups like "what does the options flow look like?" and the agent knows you mean the current ticker without you repeating it.
Privacy + model
Augur routes chat turns to DeepSeek V4-Flash and workflow-build turns to GPT-5.6 Luna, both with tool use and prompt caching. Each tool call hits a Sentimentor backend route — no third-party data broker handles your prompts.
- The model itself never executes arbitrary code — every action goes through a typed tool with backend-side validation.
- Saved reports and generated artefacts are scoped to your account; they aren't shared across users unless you explicitly export them.
- Conversation logs are not used to train the model.
Tips for better prompts
- Specify the timeframe and timeframe units. "Last 5 days" beats "recently."
- Anchor numerical thresholds where you have them. "RSI 50–75" lets the screener tool run; "mid-RSI" forces guess-work.
- State the artefact you want. "Generate a strategy" / "save this as a report" / "build a workflow" routes the agent straight to the generation tool.
- Chain in one turn. Multi-step asks ("screen then build me a strategy for the top pick") are cheaper than separate turns because the screener result stays in context.
- Use the per-tool chips for honesty. If a chip is red, take the corresponding sentence in the response with a grain of salt — the agent fell back on cached or partial data.
FAQ
Why is my answer taking several rounds?
Complex prompts naturally chain tools. The streaming bubble shows the round counter — if it's climbing, the agent is still pulling data. Most asks resolve in < 30s; research workflows can take 60–90s end to end.
A tool failed — what now?
The failing chip turns red and the agent typically explains the gap in its response ("I couldn't fetch the option chain — here's the analysis without it"). Retry the same prompt; transient backend errors usually clear within a tick.
Is the generated PineNode code production-ready?
It compiles, backtests, and respects thesafety analyzer— but you should still review it before live-running it. The editor opens with the strategy pre-loaded so you can read it line by line.