KoboldCpp v1.122 adds integrated agent with 9 tools
KoboldCpp has released version 1.122, introducing a bundled agentic harness designed as a lightweight alternative to tools like Opencode, Codex, or Claude Code. This new agent feature includes nine built-in tools and operates with a minimal system prompt of only 2k tokens. Users can enable the agent via the GUI's Admin tab or by adding the --agent flag to launch parameters. The system supports…
Key points
- KoboldCpp v1.122 ships with a bundled agent harness featuring 9 built-in tools.
- The agent requires at least 28k context and 8k generation tokens to function effectively.
- Users can enable the agent via the GUI Admin tab or the --agent launch flag.
The release also introduces three approval modes for tool calling: on, auto, and off, with a caution advised for auto-approval. To function effectively, the agent requires a minimum context of 28k tokens and a generation amount of 8k tokens, with a recommendation of at least 12GB VRAM for optimal performance. A downloadable template is available for the Qwen 3.6 35BA3B model to help users get started quickly.
Other updates in this version include the ability to specify ubatchsize independently from batchsize, similar to llama.cpp, and the removal of the pipeline parallel flag, which now auto-enables when ubatch is smaller than batchsize. The release also fixes VRAM usage issues, improves music UI, and allows for audio code generation when using LLMs for lyrics. Package maintainers should note a breaking change to the koboldcpp.sh script, which now optimizes for the target platform unless the KCPPPORTABLE=1 environment variable is set.
Koboldcpp v1.122 released
github.com · 26 September 2026Loading the full article…
This text was published by github.com. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
Comments
via GitHub Discussions