Skip to content
LVIS AI
EN|KO

Host · MCP

MCP Servers — External Tool Sets, Inside LVIS

MCP is the standard protocol that connects externally provided tool collections so they can be used inside LVIS chat. A registered server can offer more than tools — it can also publish resources (material to read) and prompts (questions prepared in advance), and LVIS surfaces each of the three in a different place.

Standard for connecting external tools
Tools · resources · prompts
Registered only after user consent

When would you use it?

  • When you want to call your company's internal database or API from LVIS chat.
  • When you want to bring a paid external service (e.g. search / translation / code analysis) into LVIS.
  • When you want to borrow a tool collection another team built, as-is.

Registration flow

  1. 01

    Discover in the Marketplace

    Check the list of available servers in the Marketplace's MCP tab. Ratings and usage stats are shown alongside each one.

  2. 02

    Press the registration deeplink

    Pressing "Register" sends a registration request to the LVIS host. You can also enter a URL directly.

  3. 03

    Review the tool list + consent

    User confirmation

    The host fetches the list of tools that server will provide and shows it to the user in advance. Risk level and scope are reviewed at this point.

  4. 04

    Joins the tool list

    After consent, that server's tools join LVIS's tool list. They can be called with natural language from chat.

What else comes in besides tools

An MCP server can offer two more things beyond tools. LVIS does not blend the three — each is surfaced in its own place and handled differently.

Resources — material the server holds

Documents, logs, records: things the server publishes as "you may read this." They arrive by two routes — the user picks one by typing @ in the composer, or the model finds and reads them itself with mcp_resource_list / mcp_resource_read.

Resource templates — material with blanks

"Give me an issue number and I'll give you that issue." Picking one from the @ list opens a host dialog to fill the blanks, and the host — not the renderer — builds the filled-in address and reads it.

Prompts — questions the server prepared

A set of questions the server prepared as "ask me this and I'll answer well." Pick one by typing / in the composer; if it takes arguments, a form appears first.

Server instructions

A note a connected server attaches saying how it wants to be used. It is passed to the model as reference material, never as instructions to follow.
What a server wrote is not what the user wrote
Resource bodies and prompt bodies are written by the server. The user chose to fetch them; the user did not write them. So LVIS does not drop that content in beside the user's own words — it wraps it in a labelled block that names its origin before handing it to the model. The host neutralizes the delimiter characters so a server cannot close that block or open a new one to pass itself off as the user, and there are length caps.
External server tools go through a prompt by default
The host sorts a tool's origin into three trust levels (built-in · plugin · external MCP) and puts external MCP servers at the lowest of the three. So on the paths that auto-allow by risk band or category, an external MCP tool is filtered out and sent to a confirmation card instead — read-only tools included.
What the user allows is remembered, though
Three documented cases skip the prompt, and all three are something the user did deliberately.① The user pressed "Always allow" on the card — that choice is stored for the same tool with the same arguments, and if the arguments change in a more dangerous direction the stored choice does not apply and the prompt returns.② The user added an allow rule for that tool in settings.③ The user switched the permission mode to "allow all."None of those three is the host promoting a tool to auto-run on its own, from usage alone.
Previous

MEMORY — What It Learned

Next

Onboarding — Getting Started Guide