SideScatter and AI
Version 0.1, effective 22 Oct 2026.
The AI assistant in SideScatter (Copilot) is optional and off until you turn it on. While it is off, a built-in offline helper answers simple requests and nothing is sent to any AI provider.
Ways to use AI
| Your own key | Hosted by Anaphase | |
|---|---|---|
| Who pays the provider | You, on your provider account | Anaphase |
| Route | Your browser, then SideScatter's server function /api/copilot, which passes the request and your key to your provider without storing either | Your browser, then SideScatter's server function, then the hosted provider |
| Whose terms apply to the provider | Yours with the provider | Anaphase's with the provider |
| Status today | Available | Limited hosted mode (see below) |
The hosted assistant has three states, shown in the app's settings:
- Not yet funded: "The hosted assistant is not yet funded. Use your own key."
- Limited: "Limited hosted mode. Free models only, with low daily limits." This is the state SideScatter is in today. Requests go through OpenRouter to free models, and OpenRouter forwards them to the model host it names for each request.
- Available: "Hosted assistant available. Requests count against your plan's AI allowance." Not switched on. It would send requests to Anthropic as Anaphase's processor, and it will not be switched on until paid plans and usage metering exist.
In the desktop preview, AI requests go from a proxy on your own computer straight to the provider or local model server you configure.
What is sent
- Your message and the conversation so far.
- On the first message, a short description of the open analysis: file name, number of events and channels, channel names and their marker labels, the current plot axes and scales, whether compensation is applied, and gate names and types.
- The results of tools the assistant runs for you. These can include population counts, percentages and medians, a compensation report and cytometer details from the file's keywords.
Per-event measurements are never sent. File names, sample names and channel labels can identify an experiment; rename them first if that matters. Do not send patient-identifiable, confidential or export-controlled material unless your organisation has approved it.
What Anaphase keeps
SideScatter's server keeps no copy of your prompts, the context or the answers. Vercel's request logs record that a request was made (time, network address, path). Anaphase does not use your content to train models.
What the provider keeps depends on the provider's own terms. In limited hosted mode those are OpenRouter's terms and the terms of the model host it routes to; free models can have different data-use terms from paid ones, so check OpenRouter's model page before sending anything sensitive. With your own key, your agreement with your provider applies.
Limits
AI output can be wrong, incomplete or out of date. AI suggestions are drafts for you to check. They are not a statistical, scientific or clinical review, and SideScatter does not mark AI output as checked. The assistant can change gates and settings in your workspace when you ask it to; review what it did before relying on it.
Hosted requests are limited per network address (30 per hour and 100 per day) with a daily limit across all users, and can stop at any time when the free models' own limits are reached. The app also stops the assistant for the rest of the session when a per-session call or estimated-cost cap is reached, as a backstop on top of your provider's own limits.
Connectors (MCP)
SideScatter has no MCP connector yet. If one is released, this page will say what it returns to your AI assistant.
Questions: paulmaciocia@gmail.com.