Compute Server

Shrimply’s optional local compute server provides model-backed transcription, text-to-speech, video segmentation, voice conversion, 3D camera tracking, and video generation. The server advertises its available devices and exact model capabilities to the editor.

Run locally

The server is a separate uv project requiring Python 3.14. From the repository root, run its locked environment with:

$ make dev-server

From the server directory, the equivalent command is:

$ uv run --locked src/main.py

Models download into their configured caches on first use. Device and memory requirements vary substantially by model; consult the model catalog before starting a large download.

Connect Shrimply

The local server listens at http://127.0.0.1:8787 by default. In Shrimply, open Preferences ‣ External. Select the local server under Inference Servers, then choose one of the compute devices reported by that server.

The Available row shows the features supplied by the selected server. If a feature is missing, it will not appear in the corresponding editor controls.

Share access

Set SHRIMPLY_SERVER_SHARE=1 to create a temporary public gradio.live URL while keeping the MessagePack API available:

$ SHRIMPLY_SERVER_SHARE=1 uv run --locked src/main.py

The public URL can invoke every compute endpoint. Share it only with trusted users, and stop the process to remove access.

Containers

The Compose configuration enables GPU access and preserves downloaded models between runs.

$ cd server
$ docker compose up --build

Compute features

See Compute Features for an overview, or open a feature directly:

Troubleshooting

Keep the server process running while Shrimply uses a compute feature. Check the selected server in Preferences ‣ External if a model is missing or a connection fails. A first request can take longer while its model downloads; later requests reuse the downloaded files.