Most “AI in a trading app” is a chatbot in a sidebar that cannot see anything you are looking at. This one can. It reads your chart, your panels, your positions and your risk limits, and it can act: add an indicator, draw a level, set an alert, write a strategy, build you a workspace, or draft an order for you to approve.
It can never place a trade on its own, and it can never move your risk limits. More on that boundary below, because it is the part worth trusting.
Opening it
An orb sits in the terminal with a line of text that changes with the page: on a chart it reads Analyze the chart of BTC/USDT, on the workflows page Build a workflow, on Discovery Find me something to trade.
Click it, or press ⌘/, and a chat window opens over the terminal. The shortcut works even while you are typing in a field, because reaching the assistant should not cost you the sentence you were in the middle of.
⌘J works too, though on the web terminal prefer ⌘/, since Chrome and Firefox on Windows and Linux take Ctrl+J for their Downloads panel before the page ever sees it. Both are rebindable in Settings → Keyboard.
Where it sits
Three placements, in Settings → Assistant:
Sidebar is the default. The orb docks in the left nav rail, so it is a tool among tools and never covers anything you are reading.
Bottom bar puts it in the bottom-right corner with the suggestion readable beside it, in a strip the terminal reserves under your panels. Same corner as floating, none of the overlap.
Floating is the loudest: the orb and its suggestion sit over your workspace, always readable, and the only placement that covers part of your layout.
The chat window itself drags anywhere by its header, and stays where you drop it across reloads. It is clamped so you can never lose it off-screen, and a reset button appears once you have moved it.
Good first questions
If you are not sure what to ask it, these all work as typed:
- “What is this chart telling me?”
- “Explain what the order book is showing right now”
- “Is Bitcoin overbought on the 4-hour?”
- “Add a 50 and 200 EMA and tell me if they have crossed”
- “Research SOL”
- “Set an alert if ETH breaks 3,200”
- “Mark my entry, stop and target on this chart”
- “Build me a workspace for trading perpetuals”
- “What did I trade this week, and how did it go?”
On an empty conversation the window offers three starters for whatever screen you are on, which is the fastest way to find out what it can do here.
Reading a run
One question can trigger a lot of work, so the chat shows you what it did rather than a spinner.
Thinking is visible. On a reasoning model you see the reasoning as it arrives, folded to “Reasoned for 12s” once the answer starts.
Tool calls collapse into one group. Open any of them to see the arguments it was given and the result it got back, which is how you check what the assistant actually read before forming an opinion. This matters: an AI answer you cannot audit is an opinion, and an AI answer with its inputs visible is a workflow.
Answers that searched the web carry their sources. If it tells you a listing is confirmed, you can see where it read that.
Code gets its own block, with a copy button and, for Python, an Open in workbench action that puts the script in your library ready to run.
Chart screenshots land in the conversation rather than being described.
Hover any answer for a copy button, which is the fastest way to get a level out of the chat and into an order ticket. Failed runs carry a Retry rather than making you retype, and every answer has a regenerate button.
While it is working
The composer never locks. Type during a run and your message waits its turn, marked Queued, and sends itself when the current answer finishes. One message waits rather than a backlog, because a second would be answered with context from before an answer you have not read.
The thread follows the answer while you are at the bottom, and lets go the moment you scroll up, because reading something further back is a decision rather than an accident. A button takes you back down and tells you which case you are in: Jump to latest, or New messages if something landed while you were away.
Minimizing does not stop anything. The orb reports what it is doing in place of the suggestion (Thinking…, Using tools…, Looking on the web…), so a collapsed assistant is never a black box. Navigating away does not stop it either: ask for a backtest, go read the order book, and the run is still going when you come back.
Your conversations
Down the left of the window, History holds every conversation, newest first. Clicking a row opens it exactly as you left it, tool activity and cards included.
Threads name themselves from your first message, then get a better title in the background. Double-click a row to rename it. Deleting asks first and is not undoable. The last 50 are kept.
Where they are stored
On your device, unless you say otherwise. Conversations live in local storage on the machine you typed them on. Nothing is uploaded by default, whether or not you have an account.
Signed in, the sidebar asks once whether to sync them. Either answer retires the question, and the switch lives in Settings → Cloud Sync afterwards.
This is the only sync domain that ships off, and the reason is worth stating: what you ask an assistant is a fuller record of what you are thinking about than a chart layout is. Uploading that should be a decision you make rather than one you discover.
With sync off, threads never leave the machine, and clearing your browser data deletes them for good. With sync on, your 25 most recent conversations travel between your signed-in devices whole.
The model sees your messages either way, because it has to answer them. Which model, and who runs it, is your choice. See AI providers.
What it can do
113 tools, and one turn can chain up to 28 of them: enough to read your chart, pull two more timeframes, write a strategy, backtest it and deploy it without coming back to you in between.
Markets. Candles, prices, order books, signals, multi-timeframe reads and pair comparisons, for any instrument on any connected exchange, not just the one on screen. It can search for instruments it does not know.
Charts. Add and configure indicators, draw lines, trendlines, rectangles and Fibonacci levels, mark entry, stop and target, change chart type and scale, add a comparison symbol, run replay, and read back what is on the chart right now, including the levels you drew yourself.
Your account. Balances, open orders, your risk limits and what today has used of them, your trade journal, your watchlists, and which venues are connected. It can see that an account exists. It can never see its keys.
Scripts and bots. Read, write and validate Python indicators and strategies, run backtests through the same engine live bots use, and deploy a strategy as a bot.
Automation. Build and edit workflows, price alerts and alert flows.
Workspaces. Read the whole panel catalogue and build you a saved layout out of it, from scratch, from a template, or from the board you are looking at. See workspaces.
Research. Web search, news, top coins, market sentiment, and a full sourced research report.
Navigation. Take you to any page, pair or workspace, and then act there.
Trading. Prepare orders for you to confirm, and cancel resting ones.
Full list in the assistant tool reference.
Three personas
Switch from the dropdown in the chat header, or Settings → Assistant.
Mentor. Explains its reasoning step by step, with analogies. Best when you are learning why a setup is a setup.
Balanced. Clear signals with enough context to judge them. The default.
Technical. Data and structure, terse. Key levels, percentages, one sentence per point. Best when you already know what you are looking at.
The persona changes how it writes, not what it can do. All the tools and all the safety rules are identical in each.
It can see what you are looking at
Every panel on screen publishes what it is showing, and the assistant is handed that on every turn. A chart reports its pair, exchange, timeframe, indicators and drawings. So “is this overbought?” is a complete question, and “add an order book to this” has something to refer to.
Pages report the record they have open, not just their own name. On Workflows the answer to “what am I looking at” is the workflow, with its name and steps. On Bots it is the deployment, with its mode and status, so an answer about “this bot” can never come back about the wrong one.
That means “tighten the stop on this” needs no follow-up question.
Prediction events go further, publishing the whole field: the question, every answer with its probability, when it resolves, and which outcome your ticket is pointed at. Ask “which of these is worth a look?” and the assistant is reading the same ladder you are.
It can point at things
When the assistant changes something you were not watching, it puts a brief glow on it: ask it to add indicators and the chart lights up; ask to see an indicator’s code and the editor does.
It is a pointing finger rather than an explanation, so the reply still tells you what changed. Navigation lights the frame by itself, because the screen changing under someone who was reading it is exactly the moment it matters whether the assistant moved you or you clicked something.
Order proposals, and the boundary
The assistant cannot place an order. What it can do is prepare one: a card in the chat with the pair, side, size and price, its reasoning above it, and paper preselected. Nothing is sent until you confirm, with the same press-and-hold gesture as the order ticket.
Ticking Don’t ask again turns that into a standing grant, either for practice trades generally or for live trades on one exchange. Grants are listed and revocable in Settings → Risk Management.
Auto-approval skips the card. It does not skip your risk guardrails, which are enforced by the code that sends orders, below the AI entirely.
This is the boundary the whole design rests on. A language model can be wrong. It can be argued into things. It can be manipulated by text it reads while researching, because it cannot always tell a web page’s instructions from yours. It should not be the thing standing between you and a blown account.
The same shape holds anywhere it builds something that could cost you money. A bot it creates is always in practice mode and switched off, and it has no tool that can arm one. Workflows and alert flows land as drafts for you to review. Simple price alerts are the one exception, because an alert nobody armed is not an alert, and it tells you when it makes one.
When a decision is genuinely yours, it asks with buttons rather than a paragraph, and waits.
What it knows about you
It sees market data your terminal already has, the panels you have open, and whatever your prompt includes. Chat history is local by default and synced only if you turn that on.
It never receives your API keys or private keys. Those live in your OS keychain or encrypted vault and are read only by the connector signing a request. There is no path from the assistant to them.
On a phone
The mobile terminal has no room for a floating window, so the assistant is one of the five tabs instead. Same conversation, same tools, same confirmation cards.
Choosing a model
The assistant runs on whichever provider you pick. Bring your own key from Groq, OpenAI, Anthropic or OpenRouter, which is free and always will be, or subscribe to hosted Pairlens Intelligence.
The reasoning loop runs on your machine either way. With a hosted model the Pairlens server is only a relay, forwarding a request and streaming a response back. It decides nothing and never sees your exchange credentials. See AI providers.
Related
- Assistant tool reference for all 113 tools
- Research reports for the long-form sourced write-up
- AI providers for keys, plans and credits
- Risk guardrails for the limits the AI cannot move