Native media engines
Image, audio, video and vector operations run through compiled C++ engines called from the agent process.
Mneme Labs / AI lab
Creative Claw.
An OpenClaw fork built for media work. Agents can inspect footage and images, find material in a local library, and call native creative engines.
How a media task moves
01
Agent
Plans a creative task and selects tools.
02
Local perception
Reads video frames or speech when enabled.
03
Native engines
Execute image, audio, video and vector operations.
04
Visual memory
Keeps searchable context around local media.
What the fork adds
Creative tools inside the agent runtime.
Local perception
Optional Ollama vision and faster-whisper speech turn frames and audio into material the agent can reason over.
Visual memory
A local SQLite vector store indexes media and supports searching a working library.
GPU coordination
A broker manages short-lived VRAM leases when local models and media operations compete for one GPU.
Local model path
Local models where they help.
The project documents optional Ollama vision using qwen2.5vl:7b and optional faster-whisper speech. The agent can still use a paid frontier model for orchestration. Local perception and native processing do not mean every part of a workflow is offline or free.
Current setup requirements
- Node 24 recommended; Node 22.19 or later is the documented minimum.
- FFmpeg on the system path for video operations.
- Native binaries and model weights provisioned separately from the repository.
- Ollama and faster-whisper only for the corresponding local perception features.
- Optional CUDA acceleration for supported neural operations on Windows.
Check the repository README for current setup and platform status before installing.
Open development
Read the implementation.
Creative Claw is public source, with its OpenClaw origin and license credited in the repo. Some inherited OpenClaw features have not been validated in this fork; the README documents that boundary.
View the GitHub repository →Learn the workflow
Start with a real media task.
Our first article explains the fork, the local model path and the hardware questions to ask before choosing a GPU. More hands-on guides will follow as the workflows are tested.
Read the Creative Claw article →