The assistant answers questions about the code in front of you and about this documentation. It is a desktop feature: the key that answers lives in the main process, and the browser build has no main process.
Where it appears
In the editor. The panel beside the source reads the strategy you have open, so a question about "this indicator" is a question about the one on screen. It writes into the editor only when you ask it to.
On a documentation page. Select a passage and an ask button appears next to it. The question carries the passage and the page you are on, which is why the answer is about this engine and not about backtesting in general.
AI Studio
Build, AI Studio is a chat beside the workspace it writes into. Describe a strategy in your own words; the answer streams into the chat, and its files land in a workspace of Build, Workspaces, the same kind you make by hand. A new chat starts a new workspace with its first file. Work in a workspace lets the AI change one you already have.
The workspace follows one rule: one file runs. Around it sit helpers in the same language, which that file imports, and the notebook. The AI keeps two notes there: README.md, what the strategy is and its rules, and journal.md, one dated entry after every change, with what was asked, the files it touched and what the engine's check said. It never rewrites a note you wrote, and it never adds a second strategy file: a different idea is a new chat.
Every change is a diff. A file the AI changed shows the lines it added and removed. Keep accepts the change, Undo puts back what the file held before. Undo on the first file of a workspace the chat made throws that workspace away. You can edit any file on the canvas yourself; it is saved into the workspace as you type.
Each strategy is checked. In Python, C#, Pine and Rust the engine compiles the file that runs with its helpers and runs it once on test bars. If that fails, the errors go back to the AI for a fix, at most twice. The check says the file works, not what it earns.
Indicators and other platforms stay drafts. An indicator is saved to your library, or onto this device. A file for MQL5, NinjaScript, EasyLanguage or another platform is downloaded. Both sit under Not in a workspace on the canvas.
The model that answers is named in the prompt box: the included AI, or the provider and model of your own key. Stop ends an answer, Write again undoes what the last answer changed and asks again, and History reopens an earlier chat.
Providers and keys
Settings, AI holds the providers, in order. The first one is the primary. The rest are the fallback: if the primary fails, the next is asked. With one provider configured there is no fallback, and the panel says so rather than letting you find out during a run.
A provider is a name, a model and a key. The model is per provider, not global, because the same name means different things at different vendors. Test sends one short request and reports what came back, so a wrong key is a sentence and not a silence.
Keys are stored through the operating system's own secret store, encrypted at rest, and they never reach the renderer. A request is made in the main process with the key attached there.
What a question costs
The allowance is counted in AI credits, which are ours and not the model provider's tokens. A provider counts words. We count requests, weighted by how much work is done for you:
| Request | Cost |
|---|---|
| A strategy written by our AI, or a long conversation with it | 5 credits |
| One short question to our AI | 3 credits |
| Your own key, saved in your account | 1 credit |
| Your own key, kept on this computer | nothing |
Bringing your own key is the cheapest on purpose. A key on this computer goes from the app straight to your provider, so we count nothing.
Where "short" stops and "long" starts is decided by the backend, not by the app: a client that classified its own requests would be a client that charges itself one credit. The thresholds are four messages in the conversation, six thousand characters across them, or an answer over 1,024 provider tokens. A strategy is always long: it carries the language reference and your canvas.
A strategy can take up to three requests. AI Studio checks every strategy it writes: it compiles it and runs it over a short stretch of test bars. If that fails, the errors go back to the AI for a fix, at most twice. Each answer is a request, so a strategy costs 5 credits when the first draft works and up to 15 when it needs both fixes.
The monthly allowance comes with the plan and expires with the month:
| Plan | AI credits a month |
|---|---|
| Free | 50 |
| Pro | 12,000 |
| Max | 60,000 |
| Enterprise | Unlimited |
Top-up packs can be bought once and never expire. The plan's credits are always spent first, and bought ones are only touched once the month's are gone. A charge that does not fit in what is left of the month comes out of the bought credits whole rather than being split. The packs, and what each one buys, are under Account and Plans, AI credits.
What to use it for, and what not to
It is good at the mechanical half: what an API call takes, why a strategy will not compile, what a statistic in the report means, how to express a rule you have already decided on.
It is not a source of trading ideas that work, and a strategy it writes is a first draft that has never been tested. Run it, look at the trades, and treat a good result on one period with the same suspicion you would treat your own.
Do not paste broker credentials, account numbers or API secrets into the assistant. The request goes to a model provider. Everything else in this app keeps secrets on your machine, and this is the one place where a paste leaves it.