KoboldCpp v1.122.1 introduces a new integrated KoboldCpp Agent, providing a lightweight agentic harness. The agent includes nine built-in tools and utilizes a minimal system prompt of 2k tokens. Users can enable this feature via the Admin tab in the GUI launcher or by using the --agent launch flag.
KoboldCpp v1.122.1 Release — Integration of KoboldCpp Agent and expanded tool support
New Features and Improvements
- Integrated Agent: Supports connection to third-party backends or OpenAI Chat Completions compatible endpoints.
- MCP Support: Users can load additional tools by providing a mcp.json file.
- Batch Size Control: Added the ability to specify
ubatchsizeindependently frombatchsize. - UI/UX Enhancements: Compacted sliders, added a
ubatchslider, and improved the music UI. - Audio Support: Allows generation of audio codes when using an LLM for lyrics via a 2-step process.
- Context Management: Increased default context size and implemented smarter handling of max output tokens when approaching the context limit.
Bug Fixes and Adjustments
- Tool Calling: Fixed issues with the tool calling parser.
- Memory and Performance: Reverted a P40 VRAM usage increase issue and added
--autoswapthresholdto prevent swapping when using small models. - Stability: Generation failures during streaming now return an error object.
- Visuals: Limited the maximum image preview size to handle pixel space diffusers and lowered the SD VAE tiling threshold to 512.
- Logging: Removed certain smartcache spam in quiet mode.
Sources
- v1.122.1 （2026-09-26）