Comparisons · buyer's guide · measured July 2026
Your agent has five ways to find a file. We ranked all five.
An AI agent on Windows reaches for one of five things when it needs a file: grep, Windows Search, fzf, the editor's own index, or a purpose-built index it calls over MCP. Judge them the way an agent actually feels them (on four axes) and only one avoids a hard tradeoff. Everything below was measured on the same 4.47-million-file machine.
TL;DR · Key takeaways
The short version. Everything below, in five lines.
Five ways, one winner. grep/ripgrep, Windows Search, fzf, the editor index, or a purpose-built index called over MCP. Only the last fills all four axes an agent feels: speed, scope, freshness and token cost.
Filenames in microseconds. 85 µs to find a name on a 4.47-million-file machine (139 µs median over 20 queries), about 412,000× faster than a VS Code workspace search and ~480,000× faster than Windows Search's 67 s.
Content in milliseconds. The whole-machine search ripgrep spent 93.8 s on comes back in 16 ms (58,625× faster), and most content queries land in ~7 to 9 ms.
One call, not seventy-one. A single indexed lookup replaces 71 grep calls and 6 m 57 s of archaeology (roughly 7,200,000× fewer calls), dropping search from ~58% of session tokens to ~0%.
Invisible and always fresh. About 44 MB at idle (less than a browser tab), and a save is searchable in ~1 ms. Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo.
The contenders
Each one is good at exactly one thing.
None of these are bad tools. Each earns its place by being the best at one job. The trouble is that an agent needs all four jobs done at once, and picks the tool it can reach, not the tool that fits.
The fallback every agent already knows how to run.
Best at: reading live bytes, never stale.
The indexer that ships with the OS.
Best at: breadth. It has already seen most folders.
Fuzzy filename matching over a piped list.
Best at: instant match, once something builds the list.
Cursor and VS Code index the repo you open.
Best at: depth inside one project, semantics included.
A background index the agent calls over MCP.
Best at: all four axes at once, the point of this article.
Scored for how an agent uses it · 4.47M files
The ranking, on one card.
Four filled pips means stronger on that axis. Read down each column and you see the tradeoff: every tool has gaps, except the last row.
Scores reflect how each tool behaves when an AI agent drives it across a whole machine, not how fast it feels on a single small repo. ripgrep keeps a perfect freshness score because it reads the live bytes every time; it loses on speed and token cost for the same reason. The editor index goes deep on one repo and stops at its boundary. Only a background index built for agents fills all four.
Machine: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. The engine is validated: crash-free with correct ranking across roughly 1,500 queries. Per-tool timing methodology in the 412,000× VS Code write-up and the ripgrep comparison.
The two you already trust
ripgrep and Cursor are the honest ceilings of two different approaches.
ANCHOR 1: THE FALLBACK
ripgrep is what your agent actually runs.
It is the best brute-force text scanner ever shipped, and its freshness is unbeatable: there is no index to go stale, because it reads the real bytes on every run. Inside a single repo it feels instant. The wall is scope: point it at a whole machine and it re-reads every byte, every query. On our 4.47M-file drive, one content search took 93.8 seconds. An agent that runs it in a loop pays that cost again and again, and streams every match back into its context window.
ANCHOR 2: THE SEMANTIC INDEX
Cursor indexes your codebase. One codebase.
Cursor builds an embedding index of the repo you open, genuinely good for “find me the code that does X” inside that project. That is a different job from an exact lookup, and it stops at the workspace boundary. Your machine is a dozen repos that ship together, plus the roughly 90% of the disk that was never in git: configs, documents, the PDF with the answer. When the question crosses a folder line, the agent falls back to the same place everyone does: ripgrep.
Semantic-in-one-repo and exact-across-the-machine are not competitors: they are different layers. Keep the editor index you like. The gap it can't cover is the whole-machine one, and that gap is where an agent spends most of its search time.
Where the ranking lands
Every tool trades one axis for another. One approach refuses to.
grep trades scope for freshness. The editor trades machine-wide reach for repo-deep understanding. Windows Search trades speed for breadth. Read the matrix as a set of pick-two's and the shape is obvious: you cannot brute-force your way to all four. You have to pay the indexing cost once, keep it fresh in the background, and hand results to the agent through the one interface it already speaks: MCP.
That is not a tool you run. It is a layer that is already warm when the agent asks.
The one that wins all four
Index once. Stay fresh. Answer in microseconds.
Interlinked Files keeps a purpose-built index of every file name and every file's contents across the whole machine: every repo, including the git-ignored files a working-directory grep never sees. A filename comes back in 85 microseconds; a content query in ~7 to 9 milliseconds. It rides the filesystem's own change signal, so a save is searchable in about a millisecond. At rest the whole engine sits around 44 MB (less than a browser tab), and the index on disk is under 1% of the drive. One signed install exposes it to 19 AI clients over MCP.
And no AI in the box. The speed comes with memory it derives itself.
The same index carries a knowledge vault, relationships it works out from your own history: files that change together, repos that share hundreds of identical files, which project is alive and which went dormant. Every edge is receipts you can check: counts and hashes, never a model's guess. It runs offline and never leaves the machine.
THE ONLY ROW THAT FILLS ALL FOUR PIPS IS THE ONE BUILT FOR AGENTS.
Common questions
The questions everyone asks first.
What's the fastest way for an AI agent to find a file on Windows?
A prebuilt, always-warm index it calls over MCP. On a 4.47-million-file machine it returns a filename in 85 µs and most content queries in ~7 to 9 ms, versus 67 s for Windows Search or 93.8 s for a whole-machine ripgrep. grep, fzf and the editor index each win one axis; only the index wins speed, scope, freshness and token cost at once.
Isn't ripgrep already fast enough?
Inside one repo, yes. Across a whole machine it re-reads every byte on every run (93.8 s for a single content search on our drive), and streams every match into the agent's context. The same query resolves in 16 ms from an index (58,625× faster), and a multi-step hunt of 71 grep calls and 6 m 57 s collapses into one call.
Doesn't the editor index in Cursor or VS Code already cover this?
Only the repo you open. Roughly 90% of a machine (other repos, configs, documents, the git-ignored files) sits outside that boundary, and when a question crosses it the agent falls back to grep. Whole-machine exact search is a different layer: keep the editor index you like and add the one that spans all of them.
How much does it cost, and does it run any AI?
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. There is no LLM in the box: no embeddings, no keys, no inference. It is an index plus a match, it runs offline, and the index never leaves your machine, so nothing is uploaded unless you publish a project or a link.
How current are the results? Does the index go stale?
It rides the filesystem's own change signal, so a saved file is searchable in about a millisecond. At rest the whole engine sits around 44 MB (less than a browser tab), and the on-disk index stays under 1% of the drive.
Give your agent the top row
Stop making it choose two of four.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing.
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 your drive.
Search whole-machine, or scope. Whole-machine when you don't know where something lives; scoped to a folder when you do. Both answer in the microsecond-to-millisecond range.
Let the vault grow. Day one it knows your git history. Every week after, it knows your machine better, automatically, with no model in the loop.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.