Your privacy choices

Sentimentor uses optional analytics cookies to understand visits and improve the product. Essential cookies remain enabled either way.

Privacy policy
Documentation

Augur · AI Research Agent

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 51 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; report drafts can be saved explicitly to /reports.
Note
Augur is launched from the animated mascot button beneath the right rail. The chat panel slides in from the right and lives over your current page, so you can keep watching a chart or workflow run while you converse.

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.

Heads up
Too vague
prompt
Look up NVDA.
Tip
Well scoped
prompt
Analyze 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:

prompt
What's the sentiment on NVDA right now?
prompt
Find me 5 momentum names with RSI 50–75 and volume above 5M, then generate a PineNode strategy for the highest-scoring one.
prompt
Compare 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.
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.
prompt
Build 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.

Note
Generated PineNode code is an editable research draft. Inspect its assumptions and validate it before relying on the result. If a requested dataset is unavailable or lacks point-in-time history, Augur should mark the missing evidence rather than infer it.

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.

Tip
Rounds cap automatically at a budget per turn. If Augur hits the cap, it summarises with what it has rather than stalling silently.

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.

Citations + sources

When Augur quotes a number or a headline, the source tool is attributed inline. For news the article URL is wired into each card and opens in a new tab. For data tools, the tool chip itself is the citation — clicking it scrolls to the rendered card with the raw response.

Tools catalogue

Augur ships with 51 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.

This counts callable tools, not upstream data sources. inspect_strategy_experiments has three actions that count as one tool, not three. search_web adds public-web discovery for product and event verification. Tools offered in a particular turn depend on the request, execution context and authorization.

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). Returns low_confidence: true when the sample window is short (30–89 days); ramps to true 52-week rank at 252+ days. Untracked tickers return iv_rank: null with a note suggesting a watchlist add.
  • get_pc_ratio_trend — put/call ratio trend over the last N days (default 5). Returns direction: '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_company_filings — dated company filing records and original filing links.
  • get_sec_financial_facts — SEC XBRL financial facts with reporting periods, units and filing provenance.
  • 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.

Reddit + X

  • get_reddit_buzz — top mentions across stock subs.
  • get_x_trending — trending tickers on X (Twitter).

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.
  • query_quiver_data — public-record research including lobbying, government contracts, 13F filings, patents and related disclosures; coverage varies by dataset.

News

  • get_market_news — a bounded sample of recent headlines per ticker, not an exhaustive product or event search.
  • search_web — discovers public sources for named products, announcements and recent events. Augur then reads relevant issuer or regulator pages with fetch_url. Empty results or provider failures mean unverified, not nonexistent. Requires a configured server-side search provider.

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 the base model has by default.

  • fetch_url — reads public web pages, including pasted URLs, with DNS and redirect checks against private/reserved destinations. Text formats only; PDF/images are unsupported. Ordinary page excerpts are capped at ~24KB; SEC excerpts at ~48KB. Up to 5 redirects within a 15-second overall timeout. Search snippets are leads, not verified full-page evidence.
  • Guest previews use the full 48-tool catalog with the same request-intent rules as signed-in research. Each preview has up to 24 tool attempts, including at most 4 web searches, within 6 tool rounds and a 110-second request deadline. Generated charts, code, workflow and report drafts can be inspected inline. Saving, activation and access to account-owned records still require an authenticated account.

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.

Strategy experiment history

inspect_strategy_experiments reads the signed-in user's retained records. It does not retrieve another user's private experiments or run a new test.

  • list — find recorded versions, including incomplete and unsuccessful validation attempts. Use the returned record IDs and next-page cursor.
  • compare — inspect two records' source and configuration changes. Performance deltas require matching test windows, data and execution assumptions.
  • outcomes — review a selected version's recorded forward observations, including open signals, losses, errors and missing fills.

These actions are read-only. They cannot save, publish, backtest, start a runner, activate a schedule or place an order. See Compare versions + follow outcomes.

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_pinenode_indicator — outputs an editable PineNode indicator draft.
  • generate_price_chart — renders historical price, volume and supported technical context from retrieved bars; not a live quote or executed trade.
  • update_candidate_set — records the conversation's candidate revision and inclusion/removal reasons; model-authored metadata, not independent verification.
  • 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 — prepares the conversation's findings as a citable report that you can save explicitly.

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.

Saved report

save_analysis_report prepares the conversation as a citable research note. Click Save on the card to add it to the report library; saved reports have public share links.

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.

prompt
Generate 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.

Tip
If the rules need real data (e.g. "sell on a dark-pool spike"), Augur typically calls the relevant data tool first, reads the recent values to anchor thresholds, then emits the strategy. You'll see both chips in the strip.

Compare versions + follow outcomes

In PineNode, open Ask Augur and choose Compare recorded versions or Follow recorded outcomes. Each opens an editable prompt; nothing is sent until you submit it. These actions inspect saved evidence, not unsaved editor changes.

prompt
Compare my recorded AAPL strategy versions. Show what changed in the rules and settings, whether the test conditions match, and what cannot be concluded. Do not change or run anything.
prompt
Review the recorded forward outcomes of my selected strategy version. Include losses, errors, open signals and missing fills. Keep hypothetical signal-close results separate from confirmed paper fills. Do not create a workflow or activate a schedule.

Augur first lists your records and asks which versions you mean when the choice is ambiguous. A comparison preserves dates, data identity, costs, sizing, engine and provenance. Missing snapshots or incompatible conditions prevent a performance-delta claim. Following outcomes reads existing observations; it does not start ongoing monitoring. If none exist, Augur says so.

Published-history integrity

Retained evidence and public visibility are different. New records default to private; Augur's history tool is owner-scoped. The public Strategy Track Record and Forward Signal Track Record show redacted projections of explicitly published records, not every user's attempts.

  • Preserve the original. New backtest evidence is append-only: revisions create new records, and a changed payload cannot reuse the same request ID. Retiring a strategy or runner preserves retained evidence; it does not erase a losing result.
  • Keep provenance visible. Client-reported, server-validation and legacy records are distinguished. Publication does not certify a result or upgrade a missing historical snapshot.
  • Do not invent execution. Hypothetical signal-close results and confirmed paper fills remain separate. Missing or ambiguous fills stay unresolved. Gross matched-price returns are not net account returns.
  • Disclose the population. Public forward history is bounded to the latest 200 published frozen records; observation and summary retrieval caps are disclosed. Selective publication is not a representative sample or proof of platform-wide performance.
Heads up
Remaining limitation: public visibility can change. There is no append-only publication/unpublication journal, permanent public snapshot or independent timestamp proof in this implementation. Registration time is not publication time. Protected database records are not a tamper-proof public archive; broker reconciliation can also update the outcome projection. These protections require the experiment-record migration and matching services.

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.

prompt
Build 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 on any paid plan, including Pro. Scheduled and on-demand runs share the same monthly allowance. 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.

prompt
Find 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).

Tip
Building a workflow is free. Only running a saved workflow counts against your monthly workflow quota (Pro:10/mo, Elite: 25/mo, Trader: 75/mo). Build as many as you want in the library and only pay for the runs.

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.

prompt
Rank 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

  1. Blind forward selection: seal an unchanged ranking request now and measure it later.
  2. Historical replay: declare a past cutoff and preserve every source used in the replay.
  3. Regime test: tag equivalent runs across bull, bear, volatile, rate and sideways conditions.
  4. Human/tool comparison: freeze the same question and score return, time, reproducibility and false positives consistently.
  5. AI ablation: preserve a simple rules baseline to test whether AI adds measurable value beyond the interface.
Heads up
A historical evidence list is a declaration, not independent proof that no later information leaked into the answer. The UI labels replays declared orunverified; only a prediction sealed near its as-of time receives the current snapshot label. Small samples and favorable regimes can still mislead. Research speed, reproducibility, and forward alpha are separate outcomes.

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. Nothing is persisted until you click Save; afterward the 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.

Note
Conversation history is kept for the panel session. Closing and reopening the panel during the same browser tab preserves your previous turns; full reloads reset.

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.

Is Augur available on every plan?

The chat panel itself is on every plan. Some upstream data tools (real-time options flow, ETF dark-pool prints) require a paid tier; the agent gracefully degrades to other sources when a gated tool is unavailable.