AI Agent

The Knovis agent helps you populate, enrich, and build environments using plain English. Tell it what you want, review what it plans to do, and let it handle the repetitive work: filling in metadata, finding images, setting up DataTags across hundreds of items at once.

What the agent can do

• Fill in DataTag fields: titles, descriptions, keywords, dates, links
• Find and set thumbnail images from the web
• Apply templates across a whole environment
• Batch-update settings across all items matching a filter
• Propose waypoints and routes based on how your content is organized
• Create items and environments from scratch given a description
• Scan a folder on your computer and generate item stubs from what it finds

Three ways to use it

1. The quick buttons at the top of the Agent panel: Check items, Find images and Fill metadata. One click runs the job on the open environment, or, when items are selected (shown as a chip above the buttons), on just those. Find images and Fill metadata need a local or cloud AI model; with Rules only they are disabled.
2. Typing in the Agent panel. Ask a question or give an instruction in your own words. The panel uses the AI provider you pick in Preferences (see the tiers below).
3. From an AI app on your computer: Claude Code, Codex or Claude Desktop. You talk to the AI app, and it works on the Knovis you have open, through the same engine as the Agent panel. See Use Knovis from Claude Code or Codex.

Asking questions

Ask the agent how something works ("which controller button plays a playlist?", "how do I change the background?", "what is my confirm threshold?") and it answers from this guide and from your Preferences, including their current values. The answer appears straight away, works with every tier including Rules only, and names the section it came from. With an AI model set up, the model rewords the answer from those same passages and nothing else.

A question never changes anything: ask "what is the title of this item?" and you get an answer, not an edit. Say it as an instruction ("set the title to …") when you want a change made.

Three tiers

The Agent panel uses one of three kinds of model behind the scenes, chosen under Preferences › AI › Provider. They're not separate features; the same panel handles all three.

Tier 1: Built-in rules

No AI model required (provider: Rules only). Tier 1 handles precise, mechanical operations using built-in logic, for things where you know exactly what you want and can describe it clearly.

Examples:

• Fetch favicons for all website items in an environment
• Select some video items, then say "set open with Knovis" so they open in the Knovis player
• Apply a specific setting to every item matching a tag or type

Tier 2: Local AI model

Tier 2 uses a small AI model that runs entirely on your machine (provider: Local model). It reads your existing items, infers what's missing, and fills it in. No data leaves your computer, and no API key is required.

The model looks at what you already have (the item label, existing tags, the target file or URL) and completes the gaps: a description, relevant keywords, a collection role. It works best for structured enrichment where the answers come from the content itself.

Examples:

• Fill in missing descriptions and keywords across a whole environment
• Find thumbnail images for items that don't have one yet
• Complete DataTag fields for a batch of items in one pass

Tier 3: Frontier AI model

Tier 3 connects the Agent panel to an external AI service: Anthropic (Claude), OpenAI, or OpenRouter. This is the most capable tier and handles tasks that need real judgment: layout decisions, generating environments from a description, understanding relationships between content, and natural-sounding metadata at scale.

Examples:

• "Build a Star Trek TNG environment organized by season, with waypoints for each one"
• "Create items for these 12 research papers and arrange them by topic"
• "Flag the most essential items in this environment and suggest a presentation route"

Tier 3 uses tokens from your API account, so each operation has a small cost. For most enrichment tasks the cost is a few cents; generating a large environment costs more.

How a typical session works

1. Open the Agent panel: View › Show Agent Panel (Ctrl+Shift+A)
2. Click a quick button, or type what you want, such as "fill in metadata for all items in this environment" or "find images for items without one"
3. The agent asks any clarifying questions it needs, then shows you what it plans to change
4. Review and approve (below); the changes are applied and saved
5. Changed your mind? Click Undo in the chat, or open the History tab and undo any earlier run

Approving changes

When the agent fills in metadata, you'll see what each item has now and what it proposes. You can:

• Accept: apply this change
• Accept + run rest: apply this one and every remaining item without reviewing each
• Skip: leave this item as it is

When it finds images, you pick from the candidates, or choose Use best + run rest to let it pick for the remaining items. A change that touches many items asks first; set how many under Preferences › Advanced › Agent › Confirm at. Stop halts a run after the current item. If the agent can't tell what you're asking, it asks rather than guesses.

Use Knovis from Claude Code or Codex

If you already use an AI app on your computer (Claude Code, Codex or Claude Desktop), it can work on Knovis directly. You describe what you want in the AI app; it reads your open environment, plans the changes, and Knovis applies them through the same engine, approvals and undo as the Agent panel.

• No API key, no token cost in Knovis. The AI app runs on your own subscription; Knovis spends nothing.
• The most capable option. These apps bring a frontier model plus their own abilities, such as reading files on your computer and searching the web, and they can work through long, multi-step jobs.
• You stay in control. Changes appear in the Agent panel for you to approve, and every run is listed in History with an Undo.

Set it up once

1. In Knovis, open Preferences › AI › AI Apps and tick Let AI apps connect to Knovis. It is off until you turn it on, and an AI app can never turn it on by itself. The status line underneath confirms Knovis is ready.
2. Click the Set up button for the AI app you use, in the same section. A line under the buttons tells you what happened and anything left to do:
   • Set up Claude Code does everything: it adds the Knovis tools to Claude Code, plus a guide that teaches Claude how to use them well. Then start a new Claude Code session.
   • Set up Codex copies the settings Codex needs. Paste them at the end of ~/.codex/config.toml and restart Codex.
   • Set up Claude Desktop copies the settings Claude Desktop needs. In Claude Desktop, open Settings › Developer › Edit Config, add them, and restart Claude Desktop.
   In Knovis 1.13.32 these are copy buttons. For Claude Code, paste the two copied lines into Claude Code one at a time, not both at once.

Then just ask

With Knovis open, talk to your AI app in plain English:

• "What's on my open Knovis board?"
• "Flag all the web videos in this environment for review"
• "Find a better image for the Dog Breed Chart"
• "Make an item for this PDF using the right template"
• "Fill in descriptions and keywords for every item that's missing them"

What you'll see in Knovis

• An approval card in the Agent panel, such as "Claude Code wants to make changes", listing what will change, with Apply and Cancel. Turn off Ask before applying changes (Preferences › AI › AI Apps) to let changes that undo can fully reverse apply straight away. Changes that can't be undone, like creating an item, always ask.
• Changes appear live on the board, and in any other window that shows the same items.
• History: each run is listed in the Agent panel's History tab, tagged with the app that made it, with an Undo.