Interlinked.

Comparisons · measured July 2026

Spotlight vs Windows Search: why desktop search is still slow.

Two search boxes a billion people use every day: ⌘-Space on a Mac, the Start menu on Windows. Both index a slice of your disk. Both answer a person typing into a box. Neither hands an AI agent a way in. Here is what each one actually indexes, why it feels slow on a real drive, and what a search layer built for agents does differently.

Spotlight
macOS only
Windows Search
67 s median find*
Your AI agent
gets neither

* median of 20 queries · Windows Search · on the 4.47M-file drive cited throughout

TL;DR · Key takeaways

·

Spotlight (macOS) and Windows Search both index only a slice of your disk and answer a person at a search box: neither exposes a call an AI agent can make, so agents fall back to walking the disk.

·

On one 4.47M-file Windows machine (Ryzen 9 9950X3D), a single file returns in 85 µs (139 µs median), about 480,000× faster than Windows Search's 67 s, and 412,000× faster than a VS Code search.

·

Whole-machine content search lands in ~7 to 9 ms, versus 93.8 s for a cold ripgrep scan (16 ms warm): a 58,625× gap.

·

A retrieval task that took 6m57s across 71 tool calls collapses to one 16 ms call: roughly 7,200,000× less waiting, and the tokens spent hunting for files drop from ~58% toward ~0%.

·

It idles at ~44 MB and a saved file is searchable in ~1 ms. Everything local is free forever, no card. Hosting starts at $5.99/mo.

The two you already know

Both were built for a human at a search box.

Neither is badly engineered. Both are decades-old products doing exactly what they were designed to do: help a person find a document they half-remember, from their own folders, by typing into a box and reading a list. The trouble starts when the drive is big, and when the thing asking the question is not a person.

macOS · Apple Spotlight

Spotlight

·

Builds a metadata index of your home folder and mounted volumes: filenames, tags, and the contents of the file types it recognizes.

·

Answers two front doors, both meant for a person: the ⌘-Space bar, and the mdfind command in a Mac terminal.

·

You can exclude folders and whole volumes; the scope is whatever Apple, or you, decided it should be.

·

It exists only on macOS. On a Windows development machine it is simply not there.

THE CATCH  Fast inside its own scope, but the scope is a slice, it is one operating system, and there is no standard interface a cross-platform agent can call.

Windows · built-in search

Windows Search

·

Indexes a default slice (your user folders and the Start menu), not the whole drive.

·

Ask for anything outside that slice and File Explorer falls back to a live directory walk.

·

Start-menu results are merged with web results, so a local lookup waits on a network round-trip.

·

Measured 67 seconds, median, to find one file on the 4.47-million-file drive below.

THE CATCH  A partial index, a web merge, and a throttled service compound into a minute-long wait on a real drive, and nothing here an agent can query directly either.

Index a slice. Answer a person. That is the shared design, and the shared ceiling.

What each one indexes, and who can reach it

Coverage is half the story. The other half is the port.

Every desktop search engine has two properties that decide whether it is useful to an agent. First, coverage: how much of the drive it has actually indexed (the filled dots below). Second, exposure: which callers it opens a door to. A human at a search box is one caller. An AI agent making a tool call is a different one, and it needs a machine interface, not a text field.

SpotlightmacOSINDEXED SCOPEHome folders + app metadataREACHABLE BYHUMAN UI ✓NO AGENT PORTvia mdfind: one person's Mac terminalWindows SearchWindowsINDEXED SCOPEUser folders by defaultREACHABLE BYHUMAN UI ✓NO AGENT PORTno standard call an agent can makeInterlinkedWindowsINDEXED SCOPEEvery file, every volumeREACHABLE BYHUMAN UI ✓AI AGENT ✓one call: any agent, any editorINDEXEDNOT INDEXEDcoverage = the dots · exposure = the two ports

Read each column top to bottom: the dot grid is the whole drive, and the filled dots are the files that engine has indexed. Below it are the two doors. Spotlight and Windows Search both open the human door only: the agent has to fall back to walking the disk itself. The third column indexes every file and opens the agent door.

Same drive · same question

Line them up, row by row.

Spotlight
macOS
Windows Search
Windows
Interlinked
Windows
Default coverageHome + metadataUser foldersEvery file, every volume
Whole-drive contentBy type, in scopePartial, throttledYes, machine-wide
Built to answerA person at ⌘-SpaceA person at StartAn agent's tool call
Interface for agentsmdfind (human CLI)None standardMCP + CLI · 19 clients
Freshness after a saveBackground, can lagHours behind, throttled~1 ms typical
Median find · 4.47M filesnot on Windows67 s139 µs
Speed vs this enginenot measured~480,000× slower

Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Windows Search and Interlinked numbers are the median of 20 queries on that drive. Spotlight is macOS-only, so it cannot be benchmarked on this Windows machine: its limits here are coverage and the missing agent interface, not raw speed inside its own scope. The 139 µs figure is the single-file median; content queries land in the ~7 to 9 ms range (exact phrases ~16 ms).

The third column, up close

A search layer built for the agent, not the search box.

Point the same two dimensions the other way. Cover every file on every volume: names and contents, every repo, plus the 90% of the disk that was never in a project folder. Keep it fresh the instant a file is saved. Then open the door the OS engines never did: a single call any agent can make.

139 µs
median find, 20 queries across 4.47M files (vs 67 s for Windows Search)
7 to 9 ms
typical content query: rare symbols ~7 ms, common words ~9 ms
~1 ms
from save to searchable, under 30 ms even in the worst case
44 MB
at rest in Task Manager, less than a browser tab
<1%
of the drive on disk, about 24 GB on a 4 TB machine
19
AI clients auto-configured in a single install

Against Windows Search on the same machine that is roughly 480,000× faster on a filename lookup: 139 microseconds instead of 67 seconds. But the speed is the smaller half of the story.

Why this matters now

The seconds aren't the point. The port is.

Suppose Windows Search were instant. It would still answer a person at the Start menu. An agent cannot type into the Start menu: it makes tool calls, and neither OS engine offers one it can use. Spotlight's closest thing is a command in a Mac terminal; Windows Search offers nothing standard at all. So every agent falls back to walking the disk on each query, which is exactly why a lookup can take minutes. A purpose-built engine opens the missing door: one call, the whole machine, the same for every agent: no per-tool wiring, no cloud, nothing leaving the box. The same index also carries short notes any agent can leave on a file or folder and the next agent reads back: no model in the loop, just an index and a match.

19 clients
Claude Code, Cursor, Copilot, Codex, Cline and more, configured by one install
1 call
the whole machine behind a single tool call, identical for every agent
0 models
no LLM, no embeddings, no keys: an index and a path match, offline

THE COMPARISONS SERIES: how the tools you already know stack up on a real drive.

Common questions

Questions, answered.

Does Interlinked replace Spotlight or Windows Search?

No. It runs alongside them. Those engines answer a person typing into a search box; Interlinked adds the layer neither has: one call any AI agent can make to search every file on every volume, offline.

How much faster is it, really?

On the same 4.47M-file machine, a filename find returns in 85 µs (139 µs median), about 480,000× faster than Windows Search's 67 s, and 412,000× faster than a VS Code search. Content search runs in ~7 to 9 ms, where a cold ripgrep scan takes 93.8 s.

Why does this matter for an AI agent specifically?

Without a search interface, an agent walks the disk on every query. One real retrieval task took 6m57s across 71 tool calls; through Interlinked it collapses to a single 16 ms call: roughly 7,200,000× less waiting, with the share of tokens spent searching dropping from ~58% toward ~0%.

Will it slow my machine down or send my files anywhere?

No. It idles around 44 MB (less than a browser tab), nothing leaves the box (no cloud, no model, no keys), and a file is searchable about 1 ms after you save it.

What does it cost?

Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo.

Do this today

Give your agent the layer the OS never built.

1

Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Cline and the rest. No JSON editing, no per-tool wiring.

2

Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB, and the index on disk stays under 1% of the drive.

3

Ask from any agent. Whole-machine when you don't know where something lives, scoped to a folder when you do. Either way the answer comes back in microseconds, machine-wide, offline.

Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.

Measured July 2026 · Ryzen 9 9950X3D · 4.47M files · Windows 11← All research