Documentation
3. Chatting With Your Chart
- Created
- Jun 5, 2026
- Updated
- Sep 20, 2026
Once a provider is configured, type a message in the assistant drawer and send it. The assistant answers with the chart as context, gathering what it needs through read tools before responding.
Chart-aware context
When you send a message, the assistant knows about the active pane: the symbol, timeframe, and series type, plus what indicators and drawings are present. To go deeper it can call read tools that run immediately and feed results back into its answer:
| Read tool | What it reads |
|---|---|
get_bars | The visible bars on a pane. |
get_drawings | The drawings on a pane. |
get_indicators | The indicators on a pane. |
capture_chart_image | A snapshot image of the chart for visual reasoning. |
get_strategy / get_backtest_results | The current strategy and its latest results. |
get_strategy_api / get_indicator_api | The Pine Script authoring references, so generated code is valid. |
get_paper_account / get_paper_positions / get_paper_orders | Your paper-trading account, open positions (with P&L and open risk), and orders — for the selected account or any archived session. |
get_paper_trades / get_paper_performance / get_paper_journal | Fills and closed trades (filterable, paged), performance statistics with breakdowns by symbol, side, time and holding period, and the account journal. See AI and Paper Trading. |
Because these read tools only observe, they run without asking — the assistant uses them to ground its answer in your actual chart.
Attachments
You can attach images (and PDFs) to a message — for example a screenshot of a setup or a page of notes — and ask the assistant to reason about them alongside the chart.
Composer controls
Below the message box:
- + opens a menu of capabilities that are off by default and remembered per browser: Search the web, Run code, and Generate images. Each shows a chip while it is on. A capability that the selected provider or model does not support is not offered.
- Think sets how much reasoning effort the model spends, from off to high. The level names and the available range differ per provider.
- The microphone dictates into the box using your browser's own speech recognition.
Cmd/Ctrl+Shift+Spacetoggles it, and Escape, typing, or sending stops it. While it is listening the hint line becomes a status line with a microphone picker, since your browser's default input is often not the one you are speaking into. - The footer shows the active provider and model, both switchable there without opening settings. Providers without a key are listed and marked rather than hidden, and the model list has a filter box.
Provider capabilities
Web search, code execution, and image generation run inside the model's own provider, not in this app, so what is available depends on who you are talking to:
| Capability | Where it runs |
|---|---|
| Web search | The provider's own search, with the cited sources listed under the reply. If a provider declines it for the selected model, the reply says so. |
| Run code | The provider's own sandbox. Claude needs 4.5 or newer and Gemini 3 or newer; Claude on Vertex and every model on Bedrock have none. |
| Generate images | OpenAI, Google, Vertex, and Bedrock. Anthropic-direct has no image model. |
Generated pictures are held for the life of the tab and are not saved. After a reload the reply shows a "not kept" placeholder rather than a broken image — the same rule attachments follow, because browser storage is a few megabytes and one image can be a large fraction of it. Download anything you want to keep before reloading.
The agent loop
The assistant works in steps: it may read context, think, propose actions, and continue, up to a small step limit per turn. The loop runs in your browser for every provider, so all five reach the same live chart state.
Above each reply is a list of what the assistant did — the tool, its argument, and the outcome — with a spinner on work still running. It collapses to a one-line summary and stays after the turn ends, so you can see what a finished answer was based on.
How replies are rendered
Replies are markdown. Formulas render, whether written as $x^2$, $$…$$, or \(…\) and \[…\]. Dollar amounts and cashtags are left as text, so a reply full of prices does not turn into broken formulas. If a reply colours a figure — a green profit, a red loss — that colour is shown.
Tips for good answers
- Be specific about the pane and timeframe you mean if several are visible.
- Ask for one outcome at a time (e.g. "add RSI 14", then "now mark divergence").
- For strategy help, describe entries and exits in plain language; the assistant can turn them into code for the backtester.
Next steps
See how the assistant's proposed changes reach your chart.