AI assistant

Chat with an LLM that builds on your canvas: notes, task lists, documents, diagrams, and shapes from a prompt. It runs a local model through Ollama or LM Studio by default (no cloud, no API keys, no per-use fees), or you can point it at a cloud provider with your own key.

First, a backend

Local (the default)

The local route needs an LLM server on your machine. Install one of these once, both free. See Set up local models ▸ AI assistant backend:

• Ollama: ollama.com/download
• LM Studio: lmstudio.ai

If nothing is connected, the AI panel shows a Setup screen that looks for both automatically and links to the downloads. A colored dot in the panel header shows connection status, green when connected.

Cloud providers (optional)

The Backend dropdown in Chat settings also lists OpenAI (ChatGPT), Claude, Gemini, Groq, xAI, DeepSeek, OpenRouter, Mistral, Together, Perplexity, and a Custom entry for any OpenAI-compatible endpoint. Pick one, paste your API key, and hit Test connection. The model list loads automatically, filtered to chat-capable models.

Each provider remembers its own key and last-used model, so switching between a local Ollama model and Claude is one dropdown change. Keys are stored encrypted on your machine (Windows DPAPI / macOS Keychain) and are only sent over HTTPS to the provider you chose.

Use it when: your GPU can’t hold a big local model and a deadline can’t wait.

The three modes

The composer has a mode dropdown:

• Suggest: the model proposes canvas actions and you click Apply to run them. Nothing changes until you approve it.
• Agent: actions run automatically as the model decides them.
• Chat: plain conversation, no canvas actions.

What it can build

In Suggest or Agent mode, ask in plain language and it creates real canvas items:

• “Make a sticky note with three logline ideas for the trailer.”
• “Add a task list for the VFX shots with statuses.”
• “Draw a flowchart of the approval process.”
• “Write a short treatment as a document node.”

Edit what’s already there

The assistant can also modify existing elements. It sees element names and editable properties, and targets items by id. Ask it to recolor, resize, rename, move, or restyle:

• “Make all the sticky notes yellow and bump their font a step.”
• “Rename the network box around the character refs to ‘Cast’.”

Organize the board

Organize cleans up the whole board or a selection into titled clusters with group boxes rather than a flat grid. It groups by content (with AI vision analysis on image boards), color, name, type, orientation, format, date added, or wired-vs-loose, and tints the boxes to match their contents. There are also bulk-style, cleanup, align, label, and scaffold commands, and multi-part instructions run every part (“group the images by color, label each group, and align the stickies on the left”).

Use it when: 200 loose items need to become labelled clusters you can present.

Settings

Open Chat settings from the gear icon in the panel header, or click the model pill. The Settings tab has Backend (Ollama / LM Studio / the cloud providers above), Model, Host (default 127.0.0.1), Port (11434 for Ollama, 1234 for LM Studio), default mode, and the system prompt. Host/Port points at an AI server on another machine on your network. For cloud backends they’re replaced by the provider’s base URL, an API key field, and Test connection.

Model is an editable dropdown listing the models installed in your backend (Refresh re-queries), or type any model tag: qwen2.5:7b-instruct, gemma3:4b, llava:7b. With Ollama selected, a Models tab offers a curated list you can download in-app.

Models that don’t work hide themselves

When a chat model fails because it is genuinely unavailable (an access, verification, or endpoint error), ReelMarkr blocklists it and drops it from the Model dropdown. Transient failures never hide a model: a timeout, dropped connection, quota, or rate-limit error leaves it in the list.

Reset hidden models sits in Chat settings, on the provider row next to Test connection. It clears the blocklist; press Refresh afterwards.

Use it when: six of forty listed models error on your tier.