
关于
A community package of the unmodified official KoboldCpp CUDA Linux build. Run local GGUF language models with a graphical launcher, a browser chat interface and an HTTP API. Model downloads have their own licenses.
Choose a model in the launcher, then open http://127.0.0.1:5001 in your browser after loading it. The CUDA and NoCUDA editions use separate ports and private data directories. This edition includes NVIDIA CUDA support and requires a compatible host NVIDIA driver for CUDA acceleration. The oldpc build is not included.
Models and settings in Documents are accessible through the launcher's file chooser. Relative downloads, configurations and databases are stored in this app's private persistent data directory. Save launcher settings as a .kcpps file and load that file on the next launch.
Command-line options can be passed after flatpak run io.github.LostRuins.koboldcpp.CUDA. For example: --skiplauncher --model /path/to/model.gguf. The server listens on loopback by default; pass --host 0.0.0.0 to make it reachable over your network.
For another model directory, grant only the directory you need: flatpak override --user --filesystem=/path/to/models:ro io.github.LostRuins.koboldcpp.CUDA. For optional microphone and audio features: flatpak override --user --socket=pulseaudio io.github.LostRuins.koboldcpp.CUDA. Undo overrides with flatpak override --user --reset io.github.LostRuins.koboldcpp.CUDA. Host toolchains and host commands are not available to agent or MCP tools inside the sandbox.
详情
权限
4 项需确认全部设备(含摄像头、USB)
网络访问
可访问互联网和本地网络
X11 窗口系统
传统显示协议;可窥探其他 X11 窗口的输入
文件:xdg-documents
进程间通信
与主机共享 IPC 命名空间