KoboldCPP v1.116 Adds MCP Support and Three New Image Models

Local AI inference got meaningfully wider this week. KoboldCPP v1.116 ships three new image generation additions alongside a structural upgrade that connects the tool to the growing MCP agent ecosystem.

On the image side, Krea 2 Turbo is the headline addition, marked as the recommended option in the release notes with suggested settings of 8 steps at 1.0 guidance. Ideogram 4 is also now supported, though it requires JSON-formatted prompts rather than plain text input. A third addition, Boogu Edit, is described in the notes as "a simple but good image editing model," enabling users to modify existing images locally rather than just generating new ones. The image generation prompt limit has also been raised to 3000 characters.

The arguably more structural change is MCP server support inside KoboldCPP's built-in Web UI. MCP has become the common connector layer for agent workflows, so this puts a local KoboldCPP instance on the same integration map as cloud-hosted models, reachable by MCP-compatible orchestrators without data leaving the local machine. The release also sets the llama.cpp jinja tool call response parser as the new default -- a change aimed at improving tool-calling reliability for agentic setups, though users with existing tool-call configurations should verify nothing breaks on update.

Several fixes accompany the new features. MTP support was corrected for Qwen and Gemma assistant models, a vision quality regression for Qwen VL mmproj was addressed, and a slowdown in Hugging Face xet file downloads was patched. On the audio side, TTS now exports MP3 via API, and there is experimental support for LTX2.3 audio-to-video with reference audio upload -- the "experimental" label on that last feature warrants caution before building any workflow dependency on it.

What the release notes do not address is hardware requirements for the new image models, whether Krea 2 Turbo and Ideogram 4 are bundled or separate downloads, and what VRAM baseline each expects. For users already running KoboldCPP for text inference, the MCP integration is probably the most immediately actionable change: it requires no additional models and opens the local instance to a wider set of agent tooling without touching the cloud.