Everything MCP Server

An MCP (Model Context Protocol) server that gives AI assistants like Claude Code instant file search capabilities via Everything by voidtools.

Everything is a lightning-fast file search engine for Windows that indexes all files and folders on NTFS drives. This MCP server exposes Everything's search as a tool that AI assistants can call, enabling them to find files across your system in milliseconds.

Prerequisites

• Windows with Everything installed and running
• Everything HTTP Server enabled (Everything > Tools > Options > HTTP Server > Enable HTTP Server)
• Node.js 18+

Installation

git clone https://github.com/Josephur/everything-mcp-server.git
cd everything-mcp-server
npm install
npm run build

Configuration

Claude Code

Run this command from the cloned repository directory:

claude mcp add --scope user everything-search -- node /path/to/everything-mcp-server/dist/index.js

This registers the MCP server globally so it's available in every Claude Code project. Replace /path/to/everything-mcp-server with the actual path where you cloned the repo.

Scope options:

• --scope user — available in all projects (recommended)
• --scope project — only available in the current project

To verify it was added:

claude mcp list

To remove it later:

claude mcp remove --scope user everything-search

CLAUDE.md Instructions

The claude mcp add command doesn't support an instructions field. To teach Claude how to prioritize this tool, add the following to your global ~/.claude/CLAUDE.md file (create it if it doesn't exist), make sure to change the port number to the port number your Everything HTTP Plugin is listening on:

## Everything Search (MCP Tool)

This machine has **Everything (voidtools)** running with an HTTP server on port 54321.
The `everything_search` MCP tool is available in every session.

### MANDATORY: Use Everything FIRST
**ALWAYS attempt `everything_search` FIRST for ANY file search** — whether by name,
path, extension, size, date, or content (`content:` prefix). This applies to every
search task, including content searches within a project. Do NOT default to Grep,
Glob, or Bash for initial searches.

### Fallback to other tools (ONLY after Everything fails or is insufficient)
If Everything returns an error (e.g. connection refused), returns no results, or
the query requires features Everything doesn't support, THEN fall back to:
- **Grep** — for complex regex content searches, or if Everything's `content:` returns nothing
- **Glob** — for precise relative-path pattern matching within the current project
- **Bash `find`** — for searches involving file permissions, symlinks, or other attributes Everything doesn't index

**Always try Everything first. Only fall back if it fails or returns insufficient results.**

### Search syntax quick reference
- `ext:py` — find by extension (multiple: `ext:ts;js`)
- `path:src\components` — match against full path
- `count:10` — limit number of results to 10
- `*.config.*` — wildcards
- `size:>10mb` — size filter
- `dm:today` / `dm:thisweek` — date modified filter
- `content:keyword` — search inside file contents
- `parent:node_modules package.json` — match parent folder
- `foo bar` — AND, `foo | bar` — OR, `!foo` — NOT
- `"exact phrase"` — literal match

This ensures Claude always prioritizes Everything over built-in search tools.

Environment Variables

Variable	Default	Description
EVERYTHING_HOST	localhost	Everything HTTP server hostname
EVERYTHING_PORT	54321	Everything HTTP server port
EVERYTHING_MAX_RESULTS	255	Max results per query — see Result Limit
PROJECT_PATH	(none)	Default project path for auto-scoping searches

Set the port to match your Everything HTTP server configuration (Everything > Tools > Options > HTTP Server).

Result Limit

To keep token usage under control, the server caps the maximum number of results returned per query to 255 by default. The AI can still request fewer results via the count parameter, but it can never exceed this cap.

If you need more results per query, increase the limit by setting EVERYTHING_MAX_RESULTS when adding the server:

claude mcp add --scope user everything-search -e EVERYTHING_MAX_RESULTS=500 -- node /path/to/everything-mcp-server/dist/index.js

Higher values return more results but consume more tokens per search. For most use cases, the default of 255 is a good balance. If you find searches are missing results, try increasing it. If you want to reduce token usage further, lower it (e.g. "100").

How It Works

The server communicates with Everything's built-in HTTP server, which returns JSON search results. The flow is:

Claude Code  -->  MCP Server (stdio)  -->  Everything HTTP Server  -->  Everything Index
                  (this project)           (port 54321)                 (NTFS volumes)

1. Claude Code calls the everything_search tool via MCP (stdio transport)
2. This server translates the request into an HTTP query to Everything's HTTP API
3. Everything searches its pre-built index and returns results as JSON
4. The server formats the results (paths, sizes, dates) and returns them to Claude

Project Scoping

By default, searches are automatically scoped to the current project folder using the project_path parameter. This prevents the AI from inadvertently searching or accessing files outside your project.

• Default behavior: The server prepends path:"<project_path>" to queries automatically
• Global search: Set global: true to search the entire system. The server will include a warning in the response reminding the AI to ask for user permission before accessing files outside the project
• Explicit path: If your query already contains a path: filter, auto-scoping is skipped

Tool: everything_search

Parameters

Parameter	Type	Default	Description
query	string	(required)	Everything search query
count	number	50	Max results to return (max 255, configurable via EVERYTHING_MAX_RESULTS env var)
offset	number	0	Result offset for pagination
sort	string	"name"	Sort by: name, path, size, date_modified
ascending	boolean	true	Sort direction
regex	boolean	false	Enable regex search mode
case_sensitive	boolean	false	Case-sensitive matching
whole_word	boolean	false	Match whole words only
match_path	boolean	false	Match against full path instead of filename
global	boolean	false	Search entire system (requires user confirmation)
project_path	string	(none)	Project folder path for auto-scoping

Search Syntax

Everything uses its own search syntax:

Syntax	Description	Example
ext:	Filter by extension	ext:py, ext:ts;js
path:	Match against full path	path:src\components
size:	Filter by file size	size:>10mb, size:1kb..100kb
dm:	Date modified filter	dm:today, dm:thisweek
dc:	Date created filter	dc:thismonth
content:	Search file contents	content:TODO (requires content indexing)
parent:	Match parent folder	parent:node_modules package.json
* ?	Wildcards	*.config.*
" "	Exact phrase	"my file.txt"
(space)	AND	foo bar
|	OR	foo | bar
!	NOT	!node_modules

Example Queries

ext:py                          # All Python files
ext:ts path:src                 # TypeScript files under any src/ folder
size:>100mb                     # Files larger than 100MB
dm:today ext:js                 # JavaScript files modified today
content:TODO ext:py             # Python files containing "TODO"
*.config.* !node_modules        # Config files, excluding node_modules

Example: Global CLAUDE.md Instructions

To teach Claude Code how to use this tool effectively, add something like this to your ~/.claude/CLAUDE.md: (make sure to change the port number to the port number your Everything HTTP Plugin is listening on)

## Everything Search (MCP Tool)

This machine has **Everything (voidtools)** running with an HTTP server on port 54321.
The `everything_search` MCP tool is available in every session.

### MANDATORY: Use Everything FIRST
**ALWAYS attempt `everything_search` FIRST for ANY file search** — whether by name,
path, extension, size, date, or content (`content:` prefix). This applies to every
search task, including content searches within a project. Do NOT default to Grep,
Glob, or Bash for initial searches.

### Fallback to other tools (ONLY after Everything fails or is insufficient)
If Everything returns an error (e.g. connection refused), returns no results, or
the query requires features Everything doesn't support, THEN fall back to:
- **Grep** — for complex regex content searches, or if Everything's `content:` returns nothing
- **Glob** — for precise relative-path pattern matching within the current project
- **Bash `find`** — for searches involving file permissions, symlinks, or other attributes Everything doesn't index

**Always try Everything first. Only fall back if it fails or returns insufficient results.**

### Search syntax quick reference
- `ext:py` — find by extension (multiple: `ext:ts;js`)
- `path:src\components` — match against full path
- `count:10` — limit number of results to 10
- `*.config.*` — wildcards
- `size:>10mb` — size filter
- `dm:today` / `dm:thisweek` — date modified filter
- `content:keyword` — search inside file contents
- `parent:node_modules package.json` — match parent folder
- `foo bar` — AND, `foo | bar` — OR, `!foo` — NOT
- `"exact phrase"` — literal match

Development

# Build
npm run build

# Run directly
npm start

# Watch mode (requires installing a file watcher)
npx tsc --watch

Project Structure

src/
  index.ts        # MCP server setup and tool handler
  everything.ts   # Everything HTTP API client
  types.ts        # TypeScript interfaces