Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

AI Integration & Assistant Control

PRISM’s AI assistant lives in the AI Assistant window: a chat interface where you talk to your show through a language model of your choice. The model can read your show — patch, groups, sequences, cues, playbacks — and propose controlled actions on it. Every write happens only when you approve it.

Setting Up the Assistant

  1. Open the AI Assistant window.
  2. Open the settings sheet and enter:
    • Endpoint — any OpenAI-compatible base URL: OpenAI, OpenRouter, Ollama, or another local model server.
    • API key / token for that endpoint.
    • Model name.
    • Kill switch — must be on for any AI write (see below).
  3. Settings are saved for the whole console, not per device.

How Chat Works

  • Reads are instant. When the assistant inspects the patch, groups, sequences, cues, or playbacks, the result appears in the chat — nothing to approve.
  • Writes are cards. When the assistant wants to change anything — set fixture values, patch a fixture, create a sequence or playback, save a cue, send a transport command — it surfaces an apply card describing the action. You Approve or Reject each one. No write executes without your tap.

Examples of things you can ask:

  • “Set the downstage wash fixtures to red.”
  • “Add a new par can to the patch and group it with the other wash fixtures.”
  • “Save the current programmer state as a new cue at the end of the main sequence with a 5-second fade, and put the sequence on playback 2.”
  • “Which cues in the intro sequence use high intensity?”
  • “Send GO on playback 1.”

Safety

Two independent gates sit between the model and your rig:

  • Kill switch. Off, and every AI write is refused (the refusal is logged). On, writes are still individual apply cards — the kill switch is a master gate, not a bypass of approval.
  • Per-action approval. Every write is an explicit Approve/Reject decision, so the operator always sees exactly what the model proposes before it happens.

Every AI tool call — arguments and result — is recorded in an audit log, viewable in the AI Assistant window.

The assistant can also read the PRISM user manual that ships inside the application (the same manual the Help button opens), so it can answer questions like “how does Flash work?” without an internet connection.

Using Your Own Model Client (aside)

The assistant’s tool backend speaks the standard Model Context Protocol (MCP), so other local MCP clients — Claude Desktop, for example — can use the same tools on your show, subject to the same kill switch and audit log. External clients cannot use the apply-card approval flow, so use the kill switch deliberately when connecting external models.

Per-AI fixture restrictions and destructive-operation blacklists are not available in the current release.