KoboldCpp v1.122 has been released, introducing an integrated Agentic harness designed as a lightweight alternative to tools like Opencode or Claude Code. This new Agent feature includes 9 built-in tools and operates with a minimal system prompt of only 2k tokens. Users can enable the Agent via the GUI launcher Admin tab or by using the --agent launch flag.
KoboldCpp v1.122 Release — Integrated Agentic Harness and Improved Performance Controls
New Features and Improvements
- Integrated Agent: Supports third-party backends and OpenAI Chat Completions compatible endpoints. It can also load additional tools via
mcp.jsonfiles. - Enhanced Control: Users can now specify
ubatchsizeindependently frombatchsize. A new--autoswapthresholdflag allows setting a size limit to prevent unnecessary swapping when using small models. - UI and Usability: Sliders in the GUI are more compact, and a new
ubatchslider has been added. The default context size has been increased. - Audio and Music: Added support for generating audio codes when using an LLM for lyrics (via a 2-step process) and improved the music UI.
- Build Script Updates: The
koboldcpp.shscript has been refactored to optimize for the target platform. Package maintainers must now includeKCPP_PORTABLE=1as an environment variable to produce builds for systems other than the build environment.
Bug Fixes
- Resolved a VRAM usage increase issue on P40 GPUs.
- Fixed several tool-calling parser issues.
- Resolved issues where generation failures during streaming failed to return an error object.
- Limited maximum image preview size to better handle pixel space diffusers.
- Synchronized upstream
sd.cpp.
Sources
- v1.122 （2026-09-26）