A position that size deserves more than a hunch.
Your own research agent — every position backed by evidence you can check.
SOURCED EVIDENCE · AUDITABLE HIT RATE · TWO WAYS TO RUN
Product flow illustration · AI-generated, research only, not investment advice
Self-Host
flagshipPut the AI subscription you already pay for to work as your analyst.
Research runs on your machine, on the subscription you already pay for. Your holdings and research never reach a model we chose or pay for (the only exceptions are Square posts and support messages, which our platform model translates or answers). One-time connector install.
Duegent Cloud
Nothing to install — we supply the compute, you just ask.
Nothing to install; the platform supplies inference. During the beta it's enabled per invite and metered in credits — contact us for a top-up.
Both modes share one research core and one webapp. The only difference is whose machine that “engine” chip runs on.
Two ways to run it
Same analysis kernel, same webapp. The only difference: where the engine runs — on your machine or ours.
| Self-Host FlagshipPut the AI subscription you already pay for to work as your analyst. | Duegent CloudNothing to install — we supply the compute, you just ask. | |
|---|---|---|
| Engine | The coding agent on your machine: Claude Code / Codex auto-detected (Kimi Code, CodeBuddy and GLM in early testing), model picked per engine, automatic switch to your backup when the primary is rate-limited | Platform-managed engine, zero setup |
| Market data | Your own free finnhub key (guided signup); platform pool as fallback | Platform market data, every number labeled source + as-of |
| Setup | Install the connector once; then the webapp detects your machine and pairs on authorize (pairing code as fallback) | Nothing to install; ready once it's enabled |
| Price | Free during beta · invite-only | Free during beta · invite-only |
Runs on the coding plans you already have
In Self-Host, Duegent doesn't sell you inference. It drives the official CLI you are already signed in to, so the subscription you already bought does the research. Credentials never leave your machine.
Claude CodeSupported
OpenAI CodexSupported
Kimi CodeEarly testing
GLM Coding PlanEarly testing
CodeBuddyEarly testing
Gemini CLISoonFrom sign-up to first briefing
Ten minutes of homework. A year of holding.
- 1
Tell Duegent what you hold
Paste holdings in plain language; Chinese names resolve too. Every change goes through a confirm card, and can be undone.
- 2
Read the daily briefing
A scheduled briefing lands in your chat: computed signals, graded risk findings, a conservative Monte Carlo band. Every number carries a source and a timestamp.
- 3
Audit the agent
Each briefing logs the signal's bull/bear stance per holding and grades it 5 and 20 trading days later. The log is yours to inspect.
Not another chatbot wrapper
Ask a generic chatbot about a stock and you get vibes. Duegent grounds every conversation in things a chatbot doesn't have:
One sharp question beats an afternoon of chat.
Answers are cheap. The question is the edge.
When you don't know what to ask, the research engine takes your rough question and guides you to a sharper, more valuable one.
- 01Knows your book
- 02Reads today's signals
- 03Learns how you ask
Deterministic signals
RSI, MACD, trend, volatility: computed from market data, reproducible, with source + as-of on every figure. Not model guesses.
A hit-rate you can audit
Every briefing logs the signal's bull/bear call on each holding and scores it 5 and 20 trading days later against the base rate. If it has been wrong, you see exactly how often.
An operator, not a talker
Recording, updating and removing holdings all happen in chat, with confirm cards and per-operation undo. A full audit trail.
Multi-market, CJK-native
US, Hong Kong, A-shares, Japan, Korea. Type a name like 腾讯 and it resolves to 0700.HK the moment you enter it. No silent guessing.
Covered markets
Holdings, pricing and technical signals cover these markets. Company events, insider trades and peer comparison come from SEC filings and are US-only; filing-grade fundamentals are US-only too, and other markets get vendor-aggregated fundamentals, labelled as a weaker tier, where available. Non-US positions price from the last close, and every report states which lanes have data and which are gaps.
Exchange wordmarks shown as tiles; official exchange logos are trademarks.
Four planes, straight from the primary source
Ask an AI and you get paraphrase. Every number Duegent gives you is stamped with its source and as-of date — for US filings, down to the accession number.
What the company itself filed
Read straight off the SEC 8-K filing index: earnings releases, executive changes, material agreements — each with its accession number and filing date, no news desk in between.
What insiders actually did
The latest Form 4 filings parsed line by line: who, what role, bought or sold, how many shares. Only open-market trades count; grants and exercises are labelled as what they are.
Annual plus latest quarter
10-K and 10-Q on two tracks: revenue, net income and EPS from the latest quarter with year-over-year; margins, ROE and valuation multiples from the latest annual report. Each one stamped with its source and date.
What retail and KOLs are saying
A six-figure decision shouldn't lose to one tweet you happened to scroll past. Reddit discussion volume and representative posts, each cited with author and link; X and KOL views require turning on web-enabled research. One-sided consensus is flagged as contrarian risk, not a signal to follow.
Plus four more lanes — technicals, macro backdrop, peer comparison, quotes. Every deep-research note opens with a coverage list telling you lane by lane what had data and what did not.
A research agent that learns you
Self-evolvingMost tools run the same prompt forever; a year in, nothing has changed. Duegent treats your usage as material: it gradually works out your habits and taste, and research starts reading like it was tuned for you.
Use
- ASKAdopted: "My MU exposure in this drawdown?"
- POSITIONNVDA 300 → 380 shares
- MEMO"Semi cycle" note updated
- DEEP0700.HK report done
Each lap = one evolution cycle. The profile only tunes delivery and emphasis — never predictions, and compliance rules are untouchable.
Every learned version sits in Settings in plain sight: inspect it, reset it, or switch the whole loop off. The profile tunes delivery and emphasis only; it never predicts, and it can never touch the compliance rules.
The methodology comes from our open-source agent-harness evolution engine PolyHarness ↗
We tested the prediction route first
We ran a pure technical signal (a three-input vote: momentum, the 50-day average, MACD) on 130–140 sector-sampled names, rolled monthly across 24 runs and averaged. At both 5 and 20 days its calls came in slightly below the base rate. That is why we never sell direction as a promise: any view here ships with its evidence, a confidence level and what would invalidate it. The numbers are published as-is, so you know we're not hiding them.
Base rate = share of all evaluated windows that moved in the called direction (up for bull, down for bear). It's the score a coin-flip caller would get.
Method: 130–140 symbols (varies by run), a 3-year daily window per run with the first 250 trading days as warm-up, then evaluated every 5 trading days (n≈13606 per run). So the conclusion can't hinge on which day the window happens to end, the same backtest is re-run at 24 monthly-spaced as-of dates (2024-10 to 2026-09); figures are the mean across runs and ± is the across-run standard deviation. The runs overlap heavily, so ± reflects end-date sensitivity, not independent sampling error. A hit = correct direction exactly 5/20 trading days later; signal stats only, no transaction costs.
Backtest results are historical statistics for informational purposes only — not a promise of future performance and not investment advice.
Pricing
- Unlimited portfolios and chat (your engine)
- An automatic briefing every weekday (Self-Host: while your connector is online)
- Full signal panel, dashboard and archive
- Everything in Free
- Research turns: multi-round digging on one question
- Planned: custom schedules, weekly report, notification channels
Nothing is billed during the beta. Features marked "Planned" are not available yet.
- Zero install, platform engine
- Everything in Pro
- Fair-use quota
Nothing is billed during the beta. Features marked "Planned" are not available yet.
Free during the invite-only beta: every invitee can use Self-Host, running on their own coding agent. The Duegent Cloud engine is enabled per invite, metered in research credits with a daily fair-use cap — contact us for a top-up when it runs out. Final pricing will be announced after the beta.
Questions people actually ask
Is this investment advice?
No. Duegent is a research tool. It computes signals, cites its sources, and, when you ask for a view, gives a lean (bullish / bearish / wait) with a confidence level and the condition that would invalidate it. It never gives price targets or tells you to buy or sell; the decision is yours. Every output is labeled AI-generated, for informational purposes only.
Will it trade my account?
Never. We don't connect to brokers, never ask for credentials, and have no order capability — none of that code exists in the product. Broker import only reads what you hand it: pasted text, a CSV/PDF parsed in your browser, or a screenshot read by the engine. Nothing is stored until you confirm the rows.
Where does my portfolio data live?
On our servers, in one database alongside your chats, reports and memory — that is what makes memory, scheduled briefings, cross-device access and hit-rate auditing work. What Self-Host changes is the inference: research runs on your own machine, so your holdings are never sent to a third-party LLM we chose or pay for (the only exceptions are Square posts and support messages, which our platform model translates or answers). Either way, you can export all of it or permanently delete your whole account whenever you want.
Which AI engines work?
Claude Code and Codex CLI are fully supported and verified. Kimi Code, CodeBuddy and the GLM Coding Plan (Claude Code pointed at z.ai) are wired in and in early testing; web research and screenshot import currently need Claude Code or Codex. Gemini CLI is in adaptation. Duegent drives the official CLI you already installed and logged into, and never touches its login credentials.
What if my engine quota runs out?
The agent says so plainly and switches to your backup engine in the order you set. If that one is capped too, it tells you the limit was hit and to retry later. It never fails silently. On Duegent Cloud, when your credits or daily fair-use cap run out, it says so too and tells you how to top up.
Is the free tier actually usable?
Yes. Unlimited portfolios, unlimited manual chat, a briefing every weekday, the full signal panel and dashboard. Self-Host inference runs on your own engine, so we have no reason to meter it.
How is this different from asking ChatGPT?
A chatbot has no memory of your holdings, no sourced and timestamped market pipeline, no reproducible math, and no record of being right or wrong. Duegent adds all four on top of the model, scores its own hit rate, and coaches your questions the other way too: its suggested questions are built from your positions and today's signals, not conversation filler.
Why does it sometimes suggest a better way to ask?
Because the ceiling on any answer is the question. Anyone can answer "how's Micron" — and it shows. "What's my position's exposure in this drawdown" is a question only your holdings and computed signals can answer. So vague questions get guided alternatives under the reply; a precise question just gets its answer, no coaching. The angles you adopt are remembered: they steer how your agent evolves.
Your positions deserve research you can check.
Sign up, say hi to the agent, and read your first research note in minutes.
Start free