Host · Chat
Tool Execution and Thinking Display
LVIS visually surfaces tool calls and what the LLM is thinking through. Every tool call branches into auto-run / confirmation card / dialog based on its risk level and category.
Thinking block
Shows what the LLM is "thinking" while composing an answer, in a light quote block. Click to collapse / expand.
Tool execution card
Tool name · a summary of the inputs used · the result, all in one card. Long results expand into a larger view via "See more."
Three sources of tools
- Built-in host tools — tools provided by LVIS itself. The most trusted.
- Plugin tools — tools registered by an installed plugin. Operate within their own domain.
- External MCP tools — tools from an external MCP server registered by the user. Treated at the lowest of the three trust levels, so they are filtered out of the risk-band auto-allow path and sent to a confirmation card.
Long-running commands, and images
- Background execution — a long-running shell command can be sent to the background. The conversation is not blocked until it finishes, and its output can be read while it runs or the command killed partway.
- Viewing an image — an image file on the PC (png · jpeg · gif · webp) can be loaded. There is a size cap, and a file that is not an image is refused. Whether it arrives as a picture depends on the model in use at that moment — on a model whose tool results are text-only, a one-line "loaded" note arrives instead of the picture.
- When the tool list gets long — as installed plugins and MCP servers accumulate, the host stops handing the model every tool at once and switches to searching for the ones it needs.
Every tool call is logged
Whether it succeeds or fails, every tool call leaves a one-line record in secure storage. External code execution is logged separately and retained longer.