Guides · the five-sign checklist
Five signs your AI agent needs a file index.
None of them look like a search problem. A session that stalls, a bill that balloons, an agent that gives up and asks you where the file is: you blame the model, the plan, the context window. On a real 4.47-million-file machine we measured what is actually happening, and all five symptoms trace to one missing piece. Here is the checklist, and how to run it on your own agent.
TL;DR · the short version
The whole piece, in five lines.
- ▸
Five unrelated-looking failures (a long pause, a bill or cap hit by lunch, dozens of tool calls per ask, an edit to a stale file, blindness across your repos) all run down to one missing piece: a file index.
- ▸
Without one, the agent walks the disk on every query. We watched a Claude Code run spend 6m 57s and 71 tool calls to find one file, with ~58% of its tokens gone to the hunt.
- ▸
With a live machine-wide index the same find returns in 16 ms and one call, and a raw filename lookup is 85 μs on a 4.47M-file machine.
- ▸
That is 412,000× faster than VS Code and ~480,000× than Windows Search; content search runs in ~7 to 9 ms where ripgrep's heaviest query took 93.8 s (58,625×).
- ▸
It idles around 44 MB and reflects a saved file in ~1 ms. Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo.
The centerpiece · symptom → diagnosis
Five symptoms. One diagnosis.
Tick the ones you recognize. They feel like five different failures (slow, expensive, chatty, stale, confused) but they run down to the same wire. The agent has no index, so every question turns into a walk across the disk. Read the checklist left to right: what you see, what it means, and the single thing that fixes all of it.
The bus is the point. You do not have five problems to triage: you have one, wearing five costumes. Give the file-hunt layer a real index and the whole column collapses: the pause, the bill, the tool-call storm, the stale edit, and the cross-repo blind spot go together, because they came from the same place.
Run the checklist · takes one session
The five signs, and how to confirm each one.
You do not have to take our word for it. Each sign comes with a test you can run inside your next agent session. If two or more land, your agent is I/O-bound on file search, waiting on the disk, not thinking.
The session stalls before it starts
Every task opens with a long, silent wait that looks like deep thought. It usually is not.
SELF-TEST · Watch the tool stream. Count the seconds between your prompt and the first real edit: most of that gap is the agent searching the disk, not reasoning about your problem.
You run out of budget early
On a metered plan the bill outruns the work; on Claude Max you hit the usage cap before lunch. Same leak, two invoices.
SELF-TEST · Check where the tokens went. Directory listings and grep output dominate: on a real session we measured ~58% of the tokens spent finding files, not answering the question.
One question, a wall of tool calls
A single ask fans out into search, list, read, grep, read again, a dozen calls before it touches your actual request.
SELF-TEST · Count the search-and-read calls in one task. Locating a file should take one call. In a measured Claude Code run it took 71 before the agent found the file it needed.
It answers from a stale copy
It edits or cites a version of a file that is no longer on disk, the one it happened to read earlier, before you changed it.
SELF-TEST · Rename or move a file mid-session, then ask about it. A grep-based agent points at the old path; a live index already knows the new one within about a millisecond of the save.
It cannot see across your repos
A folder of separate repos that ship together looks, to the agent, like a folder of strangers. The connection lives in the gap its tools cannot see.
SELF-TEST · Ask it something that spans two sibling repos. In-repo tools index one project each; the answer sits between them, where no single-repo index is looking.
The after · measured, not estimated
The same task, once the index exists.
We watched Claude Code find one file on a 4.47M-file machine, with the disk walk and then with an index behind it. Signs one, two, and three resolve in the same three numbers.
That was the whole workflow: plan, search, narrow, repeat. A single isolated lookup on the same drive was no kinder: Claude Code took 193.5 seconds where the index returned the identical result in 85 microseconds, a 2,276,000× difference. The model did not get smarter or dumber between those numbers. Only the floor under it changed.
| The searcher | The task | Time | vs Interlinked |
|---|---|---|---|
| Interlinked | one indexed lookup, same drive | 85 μs | the baseline |
| VS Code · Ctrl+P | find a file by name | 35 s | 412,000× |
| Windows Search | file query, median of 20 | 67 s | ~480,000× |
| ripgrep | content query, heaviest of the set | 93.8 s | 58,625× on the set |
| Claude Code | agent finds one file, end to end | 6 m 57 s | ~7,200,000× |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. The ripgrep row is a content query (93.8 s on the heaviest of the set → 16 ms indexed; 58,625× averaged across the five-query set); the filename rows use the 85 μs single-file lookup. Full per-tool methodology in the 412,000× write-up, the ripgrep comparison, and the Claude Code run.
Sign 02, two ways
Metered or capped, you pay for the same search.
How the wasted tokens reach you depends only on how you buy them. The waste itself is identical: the agent re-reading directory listings it will forget one turn later.
CLAUDE CODE · METERED
You pay per token, so every listing the model re-reads is billed at top-tier rates. The file-hunt tax lands straight on the invoice and stretches each task from seconds into minutes. The slowness and the bill are the same event, counted twice.
CLAUDE MAX · SUBSCRIPTION
You pay a flat rate against a usage limit, so tokens burned hunting for files are tokens off that limit. You reach the cap sooner and wait, having spent a chunk of the allowance you are paying for on ls output.
Two ways to buy the same tokens, one root cause underneath both. Whether sign 02 shows up as a fatter invoice or a limit you hit before lunch, the fix is not a different model: it is to stop making the agent walk the disk.
The fix · one index under every repo
Five symptoms. One missing layer.
A live, always-warm index of every file on the machine: names and contents, every repo, plus the ~90% of the disk that was never in git. Signs one through three are speed; four and five are reach. One index answers both. This is what closes the cross-repo blind spot, the sign no single-project tool can even see.
SAME MACHINE · SAME DRIVE · 412,000× FASTER THAN VS CODE, ~480,000× THAN WINDOWS SEARCH, 58,625× THAN RIPGREP.
Common questions · before you install
The questions we hear most.
Isn't my IDE's search or ripgrep already fast enough?
For one open project, often yes. But an agent on a 4.47M-file machine pays a fresh disk walk on every query: ripgrep's heaviest content search took 93.8 s where the index answers in ~7 to 9 ms, and a filename lookup is 85 μs, 412,000× faster than VS Code. The gap is the walk, not the tool.
Does this send my code or an AI model to the cloud?
No. There is no LLM and no inference anywhere in the box: it is an index, a hashmap, and a path/id match, running entirely offline on your machine. Nothing about your files ever leaves the disk.
Will it slow the machine down or eat memory?
At rest the engine sits around 44 MB, small enough to leave running all day. A file you save becomes searchable in about ~1 ms, so results stay fresh without a rescan.
How hard is it to connect to my agent?
One signed installer auto-configures 19 AI clients over MCP (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest) with no JSON to edit. Filenames are searchable immediately; contents fill in behind.
What does it cost?
Everything local is free forever, no card. Hosting starts at $5.99/mo.
If you ticked even two
Give your agent the index.
Install once. One signed installer auto-configures 19 AI clients (Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest) over MCP. No JSON to edit.
Let it index. Filenames are searchable immediately; contents fill in behind. At rest the engine sits around 44 MB and the index stays under 1% of the drive.
Re-run the checklist. The stall, the tool-call storm, the token drain: watch them fall off one session later. The lookup that cost 71 calls becomes one.
Ask across repos. Whole-machine when you don't know where something lives, a folder when you do. Either way it sees every repo, and the file you saved a millisecond ago.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.