Agents · the file-hunt tax · measured July 2026
Slow file search costs about $1,400 a developer, a year.
On a $200-a-month Claude Max plan, more than half the tokens your agent spends never touch your problem: they go to hunting for files. On a real 4.47-million-file machine we put a number on it: the annual bill, the part you pay in wall-clock time, and the zero-cost fix that hands it back.
TL;DR: the short version
Slow file search costs about $1,400 a year per developer on a $200/month Claude Max plan, because ~58% of an agent's tokens go to finding files, not solving them (measured, 4.47M-file machine).
You already paid for it. On a flat plan the waste hides as capacity you never use and usage limits you hit twice as fast.
The visible half is time. One measured Claude Code hunt took 6 minutes 57 seconds and 71 tool calls to find one file. With a live index: 16 milliseconds, one call.
A pre-built machine-wide index drops file-hunt tokens toward ~0% and lookups into the microsecond range: 85 μs for a single file, 139 μs median across 4.47M files.
Fixing it costs nothing: the index is free forever and hands back ~$116/month of the plan you already bought.
What the tax is
You pay two bills for file search. Neither one shows up as a line item.
The file-hunt tax is the share of an AI agent's time and token budget spent locating files and code before it does any real work. It is the most expensive thing your agent does, and the least visible, because on a flat plan there is no per-query receipt.
Anthropic's Claude Max plan lists at $200 a month for its top tier, a flat subscription, not a metered bill. So the cost of a slow lookup doesn't arrive as a charge. It arrives as two quieter things: capacity (the fraction of the plan you bought that never reaches your actual problem) and wall-clock (the minutes you sit watching the agent grind). Both are real money. Below, we price both.
(If you pay per token on the API instead of a flat plan, the same waste shows up as a cash bill that runs into the thousands. The per-lookup token math covers that model. This piece prices the flat Claude Max plan.)
Bill one · the time you can feel
One file. Six minutes, fifty-seven seconds.
We watched Claude Code look for a single file on the test machine. It made 71 tool calls and took nearly seven minutes (long enough to make a pot of coffee) before it landed. The same lookup against a live index returned in 16 milliseconds, one call. That gap is a multiplier that reads like a typo.
Not every lookup is seven minutes, but agent-heavy days run 20 to 40 of them, and the manual paths are no kinder: VS Code's Ctrl+P takes 35 seconds on this drive; Windows Search a median 67 seconds, long enough to refill your coffee, to find one file. Reclaim even ten minutes of waiting a day and that is roughly a full working week back every year. The measured speed-ups behind those numbers are in the 412,000× VS Code write-up and the ripgrep comparison (93.8 s → 16 ms on the same query).
Bill two · the capacity you can't see
The annual math, on one $200 plan.
You paid $2,400 for the year. The only question is how much of it does real work. If ~58% of the tokens go to file hunting, then in effect you bought a $2,400 plan and about $1,400 of it never reaches the problem you're solving. Here is the derivation, line by line.
| Line item | Per developer / year | Where it comes from |
|---|---|---|
| Claude Max plan (top tier) | $2,400 / yr | $200/mo × 12 · Anthropic list price |
| Share of agent tokens spent finding files | ~58% | measured, 4.47M-file machine |
| Plan capacity burned on file search | ~$1,392 / yr | $2,400 × 58% |
| Plan capacity left for real work | ~$1,008 / yr | the other 42% |
| Same tokens, with a live index | ~0% on search | measured, ~58% → ~0% |
| Plan capacity reclaimed | ~$1,392 / yr | the whole slice, back |
| Cost of the fix (Interlinked) | $0 | free forever, no card |
| Net back, year one | ~$1,392 / yr | the fix costs nothing |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. The 58% share is the token fraction measured on real agent file-hunt sessions; the dollars apply it to Anthropic's published $200/month Claude Max list price. Flat-plan framing: the same dollars you already spent, just the slice that never does real work.
Multiply it out
The whole thing, in one bar
Where your $2,400 goes. Before, and after.
Same plan, same dollars, both bars. The only change is a live index sitting under the agent, and the red slice disappears into work.
Red = plan capacity spent finding files. Blue = plan capacity doing your work. The index reclaims the red slice for free: the app, whole-machine search, and MCP cost nothing.
The payback
It costs nothing and hands the whole slice back.
The fix reclaims about $116 of plan capacity every month, and it costs nothing: the app, the index, and the MCP layer are free forever. You do not switch agents, you do not change your $200 plan; you drop a machine-wide index underneath the tools you already pay for, and the tokens that were finding files go back to solving problems. Hosting your projects is a separate, optional product starting at $5.99/mo.
AND THE INDEX ISN'T JUST SPEED. IT CARRIES A KNOWLEDGE VAULT: ON A FRESH MACHINE IT WEBBED 12 REPOS AND 8,455 GIT EVENTS INTO 26,958 RELATIONSHIP PAIRS IN 1.7 SECONDS, AND FLAGGED TWO REPOS SHARING 435 IDENTICAL FILES, UNPROMPTED.
Questions people ask
The file-hunt tax, in plain terms.
How much does slow file search actually cost a developer per year?
On a $200/month Claude Max plan, roughly $1,400, because about 58% of an agent's tokens go to locating files instead of solving the task (measured on a 4.47-million-file machine). On metered pay-as-you-go API pricing the cash bill runs higher; the per-lookup breakdown has that math.
If I'm on a flat Claude Max plan, is it really costing me money?
Yes, as capacity, not a separate invoice. The same 58% means you reach your usage limits about twice as fast and get roughly half the productive agent time you already paid for.
How does an index change the number?
A pre-built, always-fresh, machine-wide index answers a file lookup in the microsecond range (85 μs for a single file, 139 μs median across 4.47M files) with one tool call instead of dozens, which pushes file-hunt tokens toward ~0%.
What does the fix cost, and when does it pay off?
Nothing: the app, whole-machine search, and MCP are free forever, no card. It hands back about $116/month of plan capacity at zero cost. Hosting your projects is a separate, optional product starting at $5.99/mo.
Does this replace my agent or my $200 plan?
No. It is the local search layer underneath them: it makes the plan and the agent you already pay for spend their tokens on your problem instead of on directory listings.
Stop paying it
Get the slice back.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON editing. What an MCP server is, if you want the mechanics.
Let it index. Filename search works immediately; content search fills in behind it. A save is searchable in about a millisecond, and at rest the whole engine sits around 44 MB, less than a single browser tab.
Watch the tokens drop. The agent stops walking the disk on every question and asks the index instead: one call, a microsecond answer. The 58% slice collapses toward zero and the plan you bought starts doing your work.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows. For the full picture of why this layer decides how good every agent is, read the backbone of the LLM economy.