Getting started
From invite code to your first research run in ten minutes.
- 1
Register with an invite code
Duegent is in closed beta: enter an invite code at sign-up and it grants you a free token allowance, with nothing to cancel later. Verify your email and you're in.
- 2
Connect your own engine
Settings detects the Connector on your machine and pairs it in one click; if it can't see one, copy the command and run it once in a terminal. Research then runs on the coding-agent subscription you already pay for.
npm install -g @duegent/connector && duegent connect --code YOUR-PAIRING-CODE
- 3
Hand over your holdings
Three ways in: select-and-paste from your broker app; upload a CSV/PDF (parsed in your browser, the file never uploads); or drop in a screenshot of your positions and the engine reads it — mobile brokers with no copyable text included. Everything lands on a confirm card first, and every write can be undone.
Bought 320 NVDA at $120, plus 500 shares of Tencent
- 4
Ask about one holding
Ask how a holding is doing and you get a computed signal panel (price / RSI / trend / MACD / stance) plus the agent's read, every number sourced. When the question is vague, a few guided questions appear under the answer — one level deeper, built from your positions and today's signals, one tap to ask. Ask precisely and you just get the answer.
How is Micron trading right now?
- 5
Run deep research
One sentence starts it; a task card streams progress. The report covers fundamentals, technicals, news and sentiment, with industry-percentile positioning, risk findings and conservatively-framed ranges. Engines with real search can pull in the latest news, every citation sourced. The day's directional stance goes straight into the hit-rate ledger.
Run a full research pass
- 6
Come back tomorrow
Each trading morning the scheduled briefing lands in the conversation. Directional calls are logged and scored against reality 5 and 20 sessions later; over time the agent also adapts depth and emphasis to how you actually read — and that profile stays viewable, resettable, and switchable off.
What Self-Host needs
- · A signed-in coding-agent CLI on the machine running the engine (Claude Code / Codex / Kimi Code / CodeBuddy / GLM Coding Plan)
- · Live quotes are optional via your own free Finnhub key; without one, public daily closes are used and labelled with their as-of time
- · Briefing schedule is configurable; weekday mornings by default
- · Beta allowance comes from your invite code — ping us in chat when it runs low; the balance is always visible in Settings
When something goes wrong
Ordered by what you will actually see, not by subsystem — when you are stuck, what you know is what is on the screen.
The top bar says “Engine · Offline”, or nothing happens when you ask
In Self-Host mode the reasoning runs in the coding agent on your own machine. That line means the cloud cannot reach the connector there right now — usually a closed terminal, a sleeping laptop, or a changed network.
What to do
Run `duegent connect` again on that machine; the badge goes back to online within seconds. If it stays offline, the command's own output says whether the pairing lapsed or the network is blocked — paste that into feedback.
An answer stops halfway and says a limit was hit
Two different limits are easy to confuse: your OWN coding-agent subscription (Claude/Codex side), and platform research credits (Managed mode). The message says which one.
What to do
If it is your engine's limit, set a fallback engine in Settings — the agent degrades through your chosen order instead of failing silently. If it is platform credits, switch to Self-Host and use your own subscription, or contact us. The agent always names the wall it hit; it never pretends to have finished.
It cannot find my stock, or picks a different one
Resolution is a ladder: seed set, then symbol rules, then fuzzy segmentation, then the model as a fallback. When it cannot place a name it offers candidates to tap rather than guessing for you.
What to do
A symbol is always the surest input (`0700.HK`, `600519`, `2330.TW`). For genuinely ambiguous names — an A+H dual listing, say — we list every line DELIBERATELY: that is not a failure to recognise it, it is refusing to decide which one you hold. If it lands on a clearly wrong company, report it; that one is a bug.
Parts of the report say “unavailable”
That is the design, not a fault. Fundamentals, peers, material events and insider activity come from SEC filings and therefore cover US filers only; other markets get vendor-aggregated data at a clearly lower evidence tier, or an explicit gap.
What to do
A gap is a fact about our SOURCES, never a judgement about the company. “No insider filings in 90 days” and “we have no insider source for this market” are completely different statements, and the report distinguishes them. Do not read a gap as a negative signal.
Some rows were not recognised after pasting my statement
The engine reads your statement and lists what it could not place rather than inventing a symbol. Dual-listed companies land here often — nothing in the text says which line you hold.
What to do
Add the symbol for those rows and import again. We never connect to your brokerage and never ask for credentials — import does exactly one thing: parse the text or screenshot you paste.
Where do my holdings actually go
It depends on the mode, and this is the most consequential privacy choice in the product. Self-Host: inference runs on your machine under your own subscription, and your holdings are not sent to any model we chose or pay for. Managed: the research dossier goes to our AI provider under API terms that forbid training on it.
What to do
In both modes your account data (holdings, chats, research, memory) is stored on our servers — that is what makes memory, scheduled briefings, cross-device access and hit-rate auditing possible at all. All of it is exportable, and the account can be deleted permanently at any time. Details in the Privacy Policy.
Still stuck
Open a thread from Feedback in the sidebar; it reaches us directly. Include the exact wording you saw — one specific sentence locates a problem roughly ten times faster than “it doesn't work”.