Overview

What an MCP tool is

In MCP, a tool is a named, schema-described capability that an AI client can invoke on demand. The server advertises what is available, the client picks one, the server executes it, and the result flows back into the model context. Vectoralix ships three built-in tool types so you can cover the most common needs without standing up your own MCP plumbing.

The three tool types at a glance

Tool type	Purpose	Configuration complexity	External network?
File Search	Search content already uploaded to the MCP server	Low — toggle capabilities, optionally tune limits	No
Code Execute	Run user-authored JavaScript in a QuickJS sandbox	Medium — write JS plus declare typed parameters	No — sandbox has no network or filesystem
API URL Proxy	Call an external HTTP API on the model's behalf	High — URL template, mappings, headers, optional OAuth + post-processing	Yes — guarded by SSRF validator

One tool, many servers

A tool is owned by your user and organization, not by a single MCP server. A tool is linked to a server by attachment, not by copy — the same tool can be attached to many servers at once. Detach it from one server and the others are unaffected.

Tool  ──┬── attached to ──>  MCP Server A
        ├── attached to ──>  MCP Server B
        └── attached to ──>  MCP Server C

Configuration is live, not versioned

Tools are not subject to the immutable-version model that governs MCP server content. Every edit you make on a tool — code body, URL template, capability toggle, anything — takes effect immediately on every server the tool is attached to. There is no draft mode, no version pin, no rollback. If you need to stage a change, disable the tool first, edit it, then re-enable it.

How tools appear to AI clients

• Each enabled Code Execute or API URL tool registers as a single MCP tool with the AI client, named by the tool's configuration.
• A File Search tool fans out: each enabled capability (up to five) becomes its own separately named MCP tool registration.
• Disabled tools and disabled search capabilities are filtered out before registration — AI clients never see them.
• A tool whose type-specific configuration is missing is silently skipped at boot. For example, a Code Execute tool with no code body saved yet is treated as not-yet-configured rather than registered as a broken tool.

Attaching a tool to a server

Open the MCP server you want to expose the tool on, go to the Server Configuration tab, and select the tool from the tools field. Save. The tool is now live on that server. Detach the same way — clear it from the field and save. The tool itself is not deleted; only the pivot row is removed.

Picking the right tool

I want to…	Pick
let the AI search documents I uploaded to my server	File Search
add deterministic logic, formatters, calculators, or validators	Code Execute
let the AI call an external HTTP API on my behalf	API URL Proxy
verify that a configured tool actually works before pointing a real AI client at it	Testing Tools in the Playground