Chat

The home page is the chat, and you pick the model that answers: Claude, GPT or GLM. It answers questions, writes and explains code, drafts text, reads the files you attach, searches the web while it answers, and makes pictures, charts and diagrams. It cannot touch your computer from the browser, and it says so if you ask.

The chat bar

+

Add photos and files; Work on my computer (opens the download page); Connect Gmail or GitHub (plugins, coming); All plugins.

Effort

The sliders button. How hard the model thinks: Low, Medium, High, Max. See Models and effort.

Model

The pill with the model's mark. Auto, Claude Opus 5.5, Claude Fable 5.2, Claude Sonnet 5.5, Claude Sonnet 5, Claude Haiku 4.5, GPT-6 Astra; every model is listed under Models and effort.

Microphone

Dictation through your browser. Speak, and the words land in the box.

Arrow

Send. Enter sends too; Shift+Enter makes a new line.

Attachments

Press + and choose Add photos and files, or drop them into the box, or paste a screenshot. Images (JPG, PNG, WebP, GIF) and PDFs, up to four per message, 5 MB each. A picture is kept with the chat, so it shows again when you open the chat (not in a temporary chat); a PDF is sent with that message only, and saved chats keep its name, not its contents. GLM-5.3 and GLM-5.2 cannot see pictures, so PocketPhantom Noir, at the same provider, looks at them first and describes them for GLM, which can ask it about a detail it needs; the steps above the answer show it. The description is not kept either. GLM and PocketPhantom Noir cannot read PDFs, so our server takes the text out of a PDF of up to 1,000 pages and the model reads that, and more pages when it needs them; that text is not kept. A scanned PDF has no text to take out: for one of those, choose a Claude model on a paid plan. On the Free plan every model reads a PDF this way, as its text.

Answers

Answers arrive as they are written, formatted: headings, lists, tables, and code blocks with a Copy button. Links in an answer work: one to another site opens in a new tab, one to a page of ours opens in place. Where it looked something up, the sentence that came from there is followed by a small chip with the site's name, whichever model answered; press it to open the page. Under each answer: copy, and on the last one, regenerate to ask the same thing again. Stop halts an answer that is still coming.

Thinking and tools

Most models think before they answer, and some answers need a tool: a web search, or pictures made or found. While that happens, a line above the answer says what it is doing (Thinking, Searching the web, Making 3 pictures). Once it is done, the line says how long it worked and how many tools it used or, if it only thought, how long it thought. Press the line to open the steps: the thinking first, then every tool in order with what it was asked and what came back, the pages a search found among them. It stays closed until you open it and open until you close it. Copy takes the answer alone. Claude and GPT models show a summary of their thinking, which is as much as their makers share; GLM shows all of it; Claude Haiku 4.5 answers without thinking first.

Pictures

Ask for a picture in your own words and it makes it: one to four versions of the same idea at once. If you have not said how many, it asks first and offers 1 to 4 as buttons. The pictures appear above its reply; press one to see it full size, move between them with the arrows, and save or share it from there. Every picture it makes goes to your Library, even if you press Stop while it is being made. To change one, say what should be different: it makes new pictures from an adjusted description rather than editing the old ones. The Make a picture button next to + (in the + menu on a phone) works as before: one picture from exactly the words you type. The Free plan makes one picture a day, whichever way you ask; paid plans make up to 60 an hour. A temporary chat makes none.

Pictures from the internet

Ask what something looks like (a building, a bird, a painting) and it can show real pictures from Wikimedia Commons, the free media library behind Wikipedia. Each one carries a credit line saying who made it and under which licence, and, opened large, a link to the page it came from. They are fetched through our server, so your browser never contacts Wikimedia. They stay in the chat and do not go into your Library.

Price charts

• Ask for the chart or the price of one of 28 of the best-known coins (bitcoin, ether, solana and the like), a share or a pair, and the chat shows TradingView's chart of it: the coin's US-dollar pair on Coinbase, the share, or the exchange and the pair you name.
• The first TradingView chart you meet asks before it loads, because your browser then fetches it from tradingview.com, which receives your IP address, details of your device and the symbol. Always show TradingView charts loads it and, from then on, every TradingView chart by itself, in new chats and older ones, temporary ones and full screen too, in every browser and on every device you sign in on: the yes is kept on your account. Only this chart loads that one and asks again at the next. Until you choose, your browser sends nothing to TradingView. To be asked again, switch off Load TradingView charts automatically in Settings › Trading. A chart loads as it scrolls near the screen, so a long chat does not load them all at once.
• It opens on the daily chart, or on the timeframe you name, from one minute to one week, and with up to three indicators you name: RSI, MACD, Bollinger Bands, a simple or an exponential moving average (at TradingView's own length of 9, which you change in its settings on the chart), Stochastic RSI or VWAP. Name none and it shows the price and its volume alone. You can change all of it on the chart.
• Any other coin the assistant first looks up with one web search. If TradingView lists it, you get TradingView's chart of the pair the search found.
• If TradingView does not list it and the pages give its price on at least two dates, the assistant draws those prices in our own chart, with links to the pages they came from and the words Prices from the web, may be delayed or wrong. Our code checks them first: a price over five times or under a fifth of the middle price is left out, a price dated after now is refused, only pages the search returned count as sources, and no price is ever added. It is our chart described below, as a line of those prices at their dates (a dot for each while they are few): the time ranges they cover, zoom, full screen, your own levels and a high or a low among those prices marked on it, and the indicators there are prices enough for. A moving average or Bollinger Bands need one price more than their length, RSI 16 and MACD 35; in the Indicators menu, one the chart has too few prices for is dimmed and says how many it needs. An indicator you named is on when the chart opens, if the chart has it and prices enough for it. The more prices the pages give, up to 400, the more of it there is. There are no candles, no volume, no candle length and no live price, and nothing on it is fetched later.
• Looking a coin up needs web search. With it off, the assistant says so and shows no chart for that coin.
• A coin's chart of exchange prices saved in an older chat shows as TradingView's chart of the same pair.

Ask for the chart or the price of a cryptocurrency (bitcoin, ether, solana) and it shows a chart of real prices, fetched by our server when you ask. The row above the chart picks how much time it shows: 1H, 1D, 1W, 1M, 1Y or All, which is all the history there is, up to about six and a half years. Each range picks its own candle length: a minute for 1H, five minutes for 1D, an hour for 1W, four hours for 1M, a day for 1Y and a week for All. The small menu beside the row (it reads 15m, for example) changes the candle length and keeps the range; it offers only the lengths that give that range a readable number of candles, and is not there when only one does (1H and All). Candles and Line switch between candles and a line of closing prices. A chart opens on the range that fits its candle length: the one you asked for, or the one set under Charts open on in Settings › Trading, a day unless you changed it. When that range holds more candles than read well at once, it opens on the newest of them. Drag the chart to look further back. To zoom, press + or −, pinch with two fingers, or hold Ctrl and turn the mouse wheel; in full screen the wheel alone zooms. Latest takes you back to the newest candle. No indicator is on until you pick one under Indicators: moving averages (20, 50 and 200, and a 21 EMA), Bollinger bands, volume, and RSI and MACD panels; this browser remembers your choice. The full-screen button opens the chart over the page; Esc closes it. While the chart is on screen its latest price stays live, marked Live, and moves the last candle; after half an hour without you touching the page it pauses until you come back. Point at a candle, or move along it with the arrow keys, to read its open, high, low and close. The links under the chart open the coin on Coinbase or TradingView, in a new tab. The TradingView button switches the card to TradingView's own chart of the same coin at the same candle length. It loads the way every TradingView chart does (see the list above), and in it you can pick any pair and exchange and draw with TradingView's own tools. Our chart switches back to ours as you left it; the AI's lines stay on ours. Ask for the chart or the price of a share, an ETF or a fund (Apple, Nestlé), or name an exchange or a pair we do not chart ourselves (ETH/USDT on Binance, a EUR or CHF pair, a future, a currency pair or an index), and it shows a TradingView chart instead, which loads the same way. Nothing on a chart is investment advice. One answer shows at most two charts.

The chat can draw on a chart too. Your own stop-loss, entry and take-profit, from a saved strategy or from levels you named, are labelled Your stop, Your entry and Your target with their price, and your stop and target are solid lines. A strategy's stop and take-profit are measured from where it goes in: the price its entry rule names, or, on a chart of the strategy's own timeframe, the entry of a trade it is still in, and the latest price otherwise. Ask for a suggested stop-loss and take-profit and it draws them dashed and a little quieter, labelled Suggested stop and Suggested take-profit with how each was found. The code works them out from past prices: the stop just under the nearest support or two ATR (twice the average candle's range) under the price, whichever is closer but never nearer than one ATR; the take-profit at the nearest resistance at least as far above the price as the stop is below it, or at twice that distance (1:2) when there is none. As support and resistance it takes the three swing lows and highs nearest the price in the last 180 candles. The AI can add a level or a trend line of its own through prices the tools gave it, and it draws the support and resistance the analysis found and, on the strategy's own timeframe, markers where a strategy's backtest bought and sold. A suggestion is a tip, never a forecast or a reason to trade. A switch on the chart shows or hides every drawing.

Strategies

• Describing, saving, changing, renaming and deleting your strategies work as described below, in a chat or in Settings › Trading.
• To check one, ask in a chat whether your strategy triggers, or press Check in chat on it in Settings › Trading: that opens a new chat with the question typed in, for you to send. The assistant searches the web for the latest price and each value your rules read, at their timeframe (such as RSI(14) on daily candles), and our code checks every rule against those values. One check covers one coin.
• A card shows each rule with the value found, a link to the page it came from, and whether the rule is met, not met or unknown. Unknown means a value was not found (for a crossing, also the one on the candle before). Our code, not the assistant, says whether a rule is met, and it never fills a value in; the assistant is told to give only values a page states and to leave out any it does not find. Our code leaves out a value that cannot be right, such as an RSI outside 0 to 100 or an average over five times the price or under a fifth of it, and that rule is unknown too. A stop-loss or take-profit in the rules is not checked: it counts from your own entry price, which the check does not know. Values from the web may be delayed or wrong, and the card says so.
• Settings › Trading shows the last check under the strategy, kept only when it could judge both the entry and the exit rule, and only for a strategy of one coin: a check of one coin of several is not kept there.
• A check needs web search and a chat that is not temporary. With web search off, the assistant says so and checks nothing.
• Asked how a coin is doing, or to analyse it, the assistant searches the web for its recent daily closing prices, and our code works the analysis card described below out from them. It sits under our own chart of those prices, with the support and resistance the code found shaded on it, the words Prices from the web, may be delayed or wrong and links to the pages; the card says it was worked out from them. Closing prices have no intraday highs or lows and no volume, so the 52-week figure is the highest close, support and resistance are closes where the price turned, and the averages, RSI, MACD and volatility count only the latest days in a row the pages gave. A number there are too few days for says Not enough history. The assistant explains only these numbers, never a target or a direction. It needs web search, and covers coins only; asking for the chart alone still shows TradingView's.

Ask how a market is doing ("How is bitcoin doing?") and it shows an analysis beside the price chart: the price against its 50- and 200-day averages, RSI and MACD, the change over a week, a month, three months and a year, how much it swings, how far it is below its 52-week high, and the nearest support and resistance. Then it asks whether you have a strategy of your own and offers buttons: describe yours, upload it (a PDF, a screenshot or text, through +), have it suggest one to test, or keep to the analysis.

It turns your description into exact rules (when to enter, when to exit, a stop-loss or take-profit if you want one) and reads them back in plain words. Only when you say yes is the strategy saved to your account, with your own description kept word for word beside the rules; adjust it later and the new version replaces the old one under the same name, but only once you have agreed to replace it. A strategy has a timeframe: a candle a day (1D, the usual), four hours (4h) or one hour (1h). It can then backtest the rules on real prices at that timeframe, over two years of daily candles or, if you ask, up to five; about 120 days of 4-hour candles (about 100 for a coin only Coinbase lists) or about 90 days of hourly ones (about 30 for a coin only Kraken lists), less the first candles its indicators need to settle. A coin with less daily history is tested on what there is, and the result says so. It shows how the rules would have done next to simply holding: total return, return per year, the deepest drop, the share of winning trades, and every trade with why it ended. A signal on one candle's close is filled at the next candle's open, so a test only acts on what was known at the time, and every buy and every sell pays a 0.1% fee. And it can check the rules today: for each symbol, whether the entry or the exit rule is met on the latest close and, if not, how far off it is, such as RSI 41 where your rule wants under 30.

Settings › Trading lists what you saved, with its coins and timeframe, your words and the rules as sentences, to rename, change or delete. You can start a strategy there too: describe it in your own words, or begin from one of five examples to test, and it reads the rules back before anything is saved. Backtest and Check work from there as they do in the chat. A saved strategy is also in your memory, under Trading, so every chat knows it. An account keeps up to 20; they cover cryptocurrencies for now. A temporary chat shows the analysis but saves, tests and checks nothing.

Charts, diagrams and choices

When numbers read better as a picture, it draws a chart right in the answer: bars, lines, areas, a pie or a donut. Point at a value, or reach it with the keyboard, to read it; Show data adds a table of the numbers under the chart. For a process, a layout or a small illustration it draws a diagram: press it to see it large, and Show code or Download hands you the SVG. When it asks you to pick between a few options, they appear as buttons under its last answer; pressing one sends it as your reply.

Your chats

Chats are listed in the sidebar under Recent, newest first, and are saved to your account: log in on another computer and they are there, in the same order. Search finds them by title or by what was said. Rename, archive or delete a chat from its menu. Archiving takes it out of the sidebar and keeps it: Settings › Data lists the archived ones and brings them back. Deleting asks first, takes it off every computer, and cannot be undone. Chats made by a scheduled task carry a small clock.

Signed out

You can read the page signed out, but sending needs an account: the first message sends you to log in and comes back into the box afterwards.

Shortcuts

Limits

• A message is up to 32,000 characters; each request carries as much of the chat as fits, newest first (about 120,000 tokens, 24,000 on the Free plan, less on a model with a smaller window), so very long chats remember the recent part best
• 30 messages per ten minutes per account
• Four attachments per message, 5 MB each
• On the Free plan a message searches the web at most twice, reads a PDF as its text, and answers with what it has once it has read about 30,000 tokens