Choose the models behind Ape, Create and Roto.

Editing, compositing, roto, transcription and export work without a network connection. Ape and generation need their provider, unless you point Ape at a model running on your own machine.

Open Settings > Models and choose Add model. Pick a provider from the presets, paste your API key, and choose the models you want from that provider. Keys are kept in Settings and aren't shown again after you save them.

Models

Your library of configured models, with a detail view for each one.

Defaults

Which model Create offers first for each kind of job.

Ape

The conversation model Ape uses and its effort setting.

Extensions

Local plugins, the Python runtime used by model adapters, and their readiness.

Advanced

The exact request definitions behind each provider, for anyone who needs to read or write them.

To use Codex or Claude Code for Ape, Grapple can install the command-line agent, and you sign in through it with your own OpenAI or Anthropic account.

Each model states what it accepts and returns: which inputs are required, which references it takes and how many, the options it offers and their limits, and what kind of result comes back. Create builds its form from that, so the Video page for one model may offer a starting image and a duration, while another offers something different.

A model that produces images or video appears on the matching Create page. A conversation model appears in Ape's model picker. If a model can't run, Settings says what is missing, such as a key or a local file.

For a provider without a preset, Ape can do the setup. It follows a built-in method: read the provider's current documentation for the exact model name, request format, inputs, outputs and limits, reuse a compatible request definition if one is installed, and write a new one only when none fits. It then adds the model and confirms that Settings reports it as ready to run.

Ape never asks for a key in the chat. When a provider needs one, it asks you to add the provider in Settings, where the key is stored, and carries on once the connection exists.

The two roto models, EfficientTAM for tracking and ViTMatte for edges, are included with Grapple and run locally. You can switch to a compatible alternative in Settings > Models and choose it under Defaults; clearing the choice restores the included one. Transcription uses Whisper Small on your computer, and depth-based effects use a depth model you add.

Ape can also talk to a chat model you run yourself. A local LLM plugin is a small manifest that points Grapple at an OpenAI-compatible Chat Completions server on localhost, such as a llama.cpp server. The manifest contains no code, and Grapple sends its requests only to that local address.

A Create run sends the provider the prompt, settings and references you gave that run. Ape's conversation model receives what Ape reads while it works, such as the project's structure, transcripts and frames it renders to check its work, so choose a provider you are comfortable sharing that material with, or a local model. Model settings are kept by Grapple rather than inside your projects, so a project folder you share doesn't carry your keys.

Provider-backed operations are billed by the provider under your account, at their prices. You don't need a Grapple account.