AI Assistant (Oscar)

Oscar combines financial questions, sourced research and supported account actions in a conversation.

Date: 2026-09-30 Last Updated: 2026-11-14 Purpose and access Oscar combines financial questions, sourced research and supported account actions in a conversation. Open the Oscar navigation entry or /assistant; saved threads open at /assistant/c/ . Sign in to retain personal conversations and use account tools. English, Chinese, Spanish, Japanese and Korean are supported. Responsive web is not proof of identical native or voice functionality. Availability depends on your feature grants, usage limits, configured data providers and the surface you use. Research Writer and Committee have separate entitlements. Code-supported features described here are not a guarantee that every deployment has enabled them. How to use Oscar: quick practical overview Open Oscar from navigation (the Oscar bottom tab on mobile web), or go to /assistant. Type a request such as “Quick look at AAPL” and send it. Include a ticker/company and timeframe when useful. Select the lightning-bolt button in the message box for Action Prompts. On wide screens the prompts open in a side panel; on smaller screens they appear in a sheet. Some ask for a ticker before sending. Read the answer and sources, then ask a follow-up in the same conversation, such as “What are the main risks?” Where voice is available, the microphone opens Tap to talk when live-agent mode is enabled; otherwise it starts recording—tap again to stop and send the transcribed request. Microphone permission and browser/device support are required. The speaker toggle enables/disables spoken replies; a voice-input reply can also be read aloud. This is not a button to read any past answer on demand. Native apps may differ. Tools are subject to account permissions, usage limits and configuration; a prompt does not unlock them. Ask for the four-step research workflow when useful: read the recommendation, understand the fundamentals, check CNN momentum, then read news sentiment. Inputs and workflow Enter a question, ticker/company, date range and objective, or select an Action Prompt. Confirm an ambiguous company match before acting on it. Read the researching/tool activity indicators, then inspect the answer, source dates and any inline cards. A tool starting is not a successful result. Ask follow-ups or request a specific source/filing. Distinguish quotations, management statements and analyst assumptions. For account actions, inspect the returned receipt. Watchlist additions require eligible published coverage, capacity and tier access; a public-report lookup does not describe your private coverage. For research changes, select an owned workspace in a new eligible private conversation, review the before/after and citations, then use Approve selected change or Reject selected change. Ordinary prose agreement is not a substitute for the research card. Reopen the thread from Chat history, or use Copy chat for Markdown on the clipboard—not a chat PDF download. Supported requests and outputs Market data: current/historical prices, company profiles, price charts with levels or Elliott-wave annotations; partial wave counts must remain explicitly unverified. Research: read available reports and prior research, web search, filings, report verification and citations. Deep research depends on provider configuration; missing coverage is not a generated report. Market context: news, summaries, index performance, earnings, sentiment, predictions, Pattern Analysis and glossary explanations. Predictions and historical statistics are not guarantees. Personal tools: watchlists, Watch Agents, scheduled questions, feedback, supported navigation, language/theme and help. A creation receipt can be shown after an action; not every account action has a separate approval dialog. Audio: browse/generate podcasts where enabled, microphone input and spoken replies where supported. Voice sessions do not guarantee every visual card or confirmation is usable. Portfolio: portfolio/open-trade/risk reads and position-sizing calculations; paper-order execution, modification and cancellation are consequential actions, with server-enforced authorization and approval where required. These are simulations, not live brokerage orders. Multi-agent: entitled Committee requests and template runs; admin back-office tools are not general-user capabilities. Inline research: inspect workspace/evidence, capture classified notes, challenge a thesis, preview scenarios, propose thesis/model/memo changes, decisions/follow-ups, monitoring setup and lifecycle operations. See Research Writer and Thesis monitoring. Capture documents and visit notes in Oscar Open the research workspace inspector in an eligible private conversation. Select an owned folder or explicitly create one; inspect the selected company and baseline before saving. Type a note or choose Attach PDF or Word file. The file picker accepts .pdf and .docx (one file per request, maximum 20 MB); legacy .doc is not the documented input. Set Source classification to management statement, analyst observation, assumption or public evidence. Mark Restricted material (keep private) when applicable. A classification label is not independent fact-checking. Save/attach and wait for the saved notice and evidence receipt. The operation is bound to the selected folder. Saving evidence does not silently alter the thesis, model or memo. Read the captured passage, then request a sourced change if needed. Check its before/after values and citations separately before approval. A failed upload or pending capture is not saved evidence. The inspector also shows baseline differences and added sources, links to the selected Research Writer folder, and allows inspecting persisted operations. Source text is research material, not authorization to act. Restricted captures can prevent export/publication even after an otherwise valid edit. Practical research and market workflows Ask “Show AAPL historical prices for the last year, with support and resistance.” Specify the period and requested levels; check the returned date range and whether an annotation is verified. Ask “Read the latest report and show both model views if they disagree.” Existing reports can include primary and second-model opinions; neither should silently replace the other. Ask “What does the latest filing say about demand risk?” Filing content lookup supports SEC filings and Japanese EDINET/TDnet sources where available. Ask for passages rather than treating a link or filing date as read content. Japanese cold lookups can be slow; avoid repeatedly submitting the same request. Ask for a prediction refresh explicitly to bypass its cache. A forecast horizon is distinct from the historical lookback window; refreshing does not make the prediction certain. Ask “Calculate position size from my risk budget and stop.” Supply assumptions rather than allowing invented capital or stop values. A sizing calculation is not an order. Use Settings → Notifications to inspect, pause/resume and delete recurring questions and read their recent results. Creating/editing the question and requesting immediate execution remain chat actions. Which actions actually require approval? | Surface/action | Actual boundary | | --- | --- | | Research evidence capture | Explicit Save/Attach saves evidence; it is not a proposal to change the thesis. | | Research thesis/model/memo, decisions/follow-ups and monitoring configuration | Oscar prepares a durable proposal. Inspect and approve/reject the selected operation in the research card; a chat reply does not apply it. | | Research export/review/publish | Separate content-bound confirmation, expiring after five minutes; an applied receipt is required. Review means snapshot creation, not sign-off. | | Builder template paper execution/cancellation/modification, report regeneration, run-schedule-now | The template executor pauses for server-issued, single-use approval bound to the exact action arguments. Template paper trades use Builder Sandbox rather than implicitly changing the main Discovery simulation. | | General account tools | Do not infer a universal approval dialog from template behavior. Watch-rule/schedule creation can return a post-action receipt; Settings schedule deletion is immediate. | The paper-trade handlers and the template approval gate are different layers. The verified hard approval boundary above is the template executor; this documentation does not claim every general-chat or headless invocation has the same card gate. Never interpret a read, calculation or pending approval as an executed trade. Boundaries and states Research scenario previews calculate before/after values only: they do not save assumptions or alter a report. Research review preparation creates a snapshot, not workflow submission or approval. Export/publication requires its own content-bound confirmation; pending cards are not downloads or published reports. Restricted captures can block outbound export/publication. Expired or stale proposals require a fresh preview. The embedded Research Writer copilot is a narrower draft-only surface. Scheduled/headless runs exclude research operations, schedule management, navigation, language/theme changes and several interactive watchlist/audio tools; do not assume the full chat toolbox runs unattended. Voice can require returning to the visual interface to inspect/approve research. An empty conversation offers prompts, not completed research. Loading/tool activity can precede partial results. Access denied, quota exhaustion, unavailable sources and failed operations must be treated as errors, not success; check persisted state before retrying a consequential action. Buy/Sell/Hold access remains subject to compliance.