AI Assistant

The Agent module is a chat interface to an AI assistant for strategy development. The assistant can analyze your strategies and backtests, suggest improvements, and answer questions about how they work.

What the assistant can see

The assistant can read your saved strategies and indicators, the strategy you currently have open in the Editor, and your most recent Backtest Lab run. When you ask whether a strategy works, it reads the run's deflated Sharpe probability — the probability that the observed Sharpe exceeds the expected maximum across how many parameter combinations you tried to find it — rather than the raw number, because a high Sharpe found across hundreds of combinations is usually selection bias rather than edge.

Proposed edits

The assistant can propose a change to a strategy. It never writes to your code directly: a proposal opens a review pane showing the current source beside the proposed source, with any compile errors listed. Nothing changes until you click Apply.

In the main assistant panel, proposals can be reviewed and applied without leaving the chat; in Agent tabs within other modules, the assistant will decline a proposal and ask you to use that panel instead.

Saved chats

Every conversation is saved as you go, after each reply, and comes back where you left it: an Agent tab reopens with the chat it held after a reload, and the assistant pane keeps its chat across workspaces. Two controls beside the chat box do the rest — New chat starts an empty one, and History lists every saved chat, pinned first then newest, with a search box that reads titles, folder names and the messages themselves. From a chat's … you can rename it, pin it, move it to a folder, export it as a Markdown file, or delete it. A chat's title is only for finding it; it comes from your first message until you rename it, and the Agent tab's own title never changes with it.

In the sidebar workspace mode the same chats appear in the Files pane under Agent Chats, where you can make folders, drag chats between them, rename and delete — the way strategies and watchlists work there.

Where chats go is Settings → Storage → Save Chats To:

• Vectaris — the Early Supporter option: your chats are stored on our servers and follow you across devices. A chat that has not reached the server yet says Not synced yet in History and goes up when the connection returns. If your subscription ends, what is already there stays readable for 90 days and is copied into your browser; new chats stay in the browser.
• Local — the default for free accounts: chats are kept in this browser only.
• Don't save — chats last for the session.

Caps keep storage bounded: 500 chats and 100 MB synced, 200 chats and 50 MB in a browser. Past a cap the oldest unpinned chat is removed and History says so. The model's view of a long chat is its last twenty turns; the transcript you read is kept whole.

Bringing your own key

Vectaris does not ship a model key. You supply a key from Anthropic, OpenAI, or Google, and your usage is billed by that provider rather than by us — which is why there is no cap on how much you use the assistant.

Several providers, one chat

You can keep a key for each provider at once. AI Settings (the gear on the assistant pane, or the same entry in any Agent tab) lists the three providers with whether a key is saved for each; pick a row to add or replace its key and choose its model, and Save stores every key you typed in that visit. The selected row is the provider the assistant uses.

Once more than one is set up, the model menu beside the chat box — it names the provider and model answering you — switches between them in one click. A switch applies to the assistant pane and every open Agent tab together, keeps the conversation you were having, and comes back to the model you last chose for that provider. Remove the active provider's key and the assistant moves to another one you have set up rather than stopping.

Your key goes to the Vectaris servers once, when you save it, and is never sent back to your browser: the servers add it to each request they relay to the provider. Where it is kept is your choice, under Settings → Storage → Save Connections To, the same setting that governs your broker connections:

• Vectaris — sealed on our servers and synced to your account, so it is there on your next device.
• Don't save — held in a Vectaris server's memory only, until you sign out, it sits unused for eight hours, or that server restarts. Nothing is written to storage, and you enter it again afterwards.

There is no option that keeps the key in your browser.

The Backtest Lab works fully without a key; the assistant is the only surface that needs one.