LM Studio adds agent safety controls in latest release; company publishes no compliance certification or liability terms

LM Studio, a desktop application for running AI models and agents locally, shipped a round of governance-oriented changes in its latest release, according to a changelog reviewed by AI Dispatch. The update adds "permission prompts for destructive file deletion operations," the changelog states — a runtime check before an agent can delete files on a user's machine.

The release also changes how LM Studio connects to external tools. It ships a "clearer button to complete authentication for OAuth MCP servers," addressing the Model Context Protocol, the standard several AI vendors use to let models call outside tools and data sources. The software already lets multiple agents work within one session, and the changelog notes "more consistent progress indicators while tools and subagents are running" — evidence of that multi-agent orchestration inside a single session.

Other changes in the release include automatic context-window sizing for local MLX-format models, "based on available memory," and improved rendering of Hugging Face model cards inside the app: "better Hugging Face model card rendering," per the changelog. The update also fixes a bug in the app's built-in API server, whose start and stop controls were not surfacing errors in settings.

The permission-prompt change is the kind of enforced-at-run-time limit AI Dispatch's trust-gap rubric scores under runtime governance. But it stands without the documentation that typically surrounds it. LM Studio's public pages, reviewed by AI Dispatch, name no compliance certification, no service-level agreement, no independent third-party evaluation of its security practices, and no liability cap for when an agent's action causes harm. The company has not published an audit-trail or incident-reporting policy either.

LM Studio is free, open-weight-model-focused software aimed at individual developers rather than enterprise deployments, which may explain the absence of formal assurance terms most enterprise vendors publish. For a team weighing it against a hosted alternative on governance grounds, that documentation currently does not exist to compare.