28.69×
Enter the workspace sooner
First file in the whole repository · versus depth=1
ScorpioFS11.824s
Git339.282s
n=1 each · single observation
FUSE workspaceOn-demand accessAgent-first
A million files. Start with what you need.
An on-demand workspace for AI agents and large monorepos. Mount a remote repository as a local filesystem, read the files the work needs, and continue from there.
Million-file run · first file usable in 11.824 s · server already prepared · n=1
/project/svc0225,000 files in scope
The subrepository view stays on the path the task needs. Content is read when something asks for it.
Interactive diagram only. It does not represent measured bytes or time.
02 — Measured, not imagined
Start sooner. See the change sooner.
From the first usable file in a million-file repository, to a new workspace and a complete patch. One cloud run answers whether the wait is worth it.
See every measured advantageALIYUN ACS / 2026-10-21One run · each result uses the faster Git baseline for that operation
4.61×
Create the next workspace sooner
Reused base · versus a native Git worktree
ScorpioFS9.162s
Git42.244s
n=3 each · median
106.13×
Get a complete diff sooner
25k subrepository · one file · first response
ScorpioFS1.350ms
Git143.244ms
n=1 each · single observation
First usable includes the first target-file read and its SHA-256 check. The server was already prepared; this is not a full download. Explicit server preparation waited 258.59 seconds and is listed apart from the client speedup. New workspaces reuse their own prepared base. Diff times are ScorpioFS’s native interface returning a complete patch. See the measurement definition and full scope
- repository files
- 1,000,000
- logical file data · decimal
- 8.32 GB
- simulated commits
- 10
- runner and backend, each
- 4 CPU / 16 GiB
03 — Less setup. More context.
A lighter entry for the next piece of work.
A developer needs a directory. An agent needs a workspace. ScorpioFS connects a remote monorepo to the local machine and brings code as the work asks for it.
Mount it, then work.
Map a remote repository path to a local workspace. First ready means the target file can be read and its contents check out.
Remote path → local workspace
Open only the code you need.
Choose a subdirectory or the whole monorepo. File contents arrive when they are accessed, so the workspace does not materialize every source file first.
One directory, one service, or the whole repository
Make the change visible sooner.
A native interface returns a dirty check and a complete patch. In the measured subrepositories, less time goes to waiting on diff, and the result can be checked on its own.
Changes → a complete, verified patch
04 — Workspaces for agents
One base. Many starts.
Creating workspaces over and over repeats the cost of getting ready. ScorpioFS reuses a prepared base and gives each task its own workspace, so the next one can start sooner.
Learn about ScorpioFS9.16sNew million-file workspace
n=3 each · median · 4.61× faster than Git worktree
n=3 each · median · 4.61× faster than Git worktree
Shared base / Independent workspacesScorpioFS v3
Prepared repository base
01workspace / a
02workspace / b
03workspace / c
Workspace boundaryIndependent
File accessOn demand
Tool entryLocal filesystem
A diagram of independent workspaces, not a concurrency benchmark. 9.16 seconds is the measured time to create those workspaces one after another.
05 — The diff, without the wait
Spend the time on what changed.
One 25,000-file subrepository and three change sizes. Receive the full patch, check SHA-256, then verify forward apply and reverse restore independently.
ScorpioFS / Verified complete patch
106.13×
Speedup over the faster Git baseline
1 file change · first response · n=1 each. Timing includes receiving the complete patch and computing its digest. Apply and reverse are checked independently after the timer.
The first response and the repeat response with cache retained are shown separately. The retained phase uses the median of n=2. Diff advantages for the 250-file subrepository are in the full table. See the full table
Finding whether anything changed can be quick too.
A 25k subrepository with 100 changed files. This compares only the boolean “is it dirty?” answer with Git diff --quiet.
1.528ms6.75× · median of n=3
06 — Remote content. Local interface.
Your tools. A familiar file entry.
Editors, terminals, and agents read code through a local directory. ScorpioFS uses FUSE to connect a remote monorepo path to the workspace. Mega2 provides the repository backend and object storage.
Editor
Terminal
AI Agent
FUSE / Local filesystemScorpioFSOn-demand access · separate workspaces
Mega2Monorepo backend · object storage
Architecture diagram. It does not mean every tool has completed compatibility testing.
07 — Every advantage, accounted for
Every advantage has the raw record behind it.
All 13 advantages from this run. Each row keeps its scope, cache phase, faster Git baseline, sample count, and statistic. Times are milliseconds.
Download the advantage data13 / 13 records
| Scope | Operation | Phase | ScorpioFS / ms | Git / ms | Speedup | Git baseline | n: Git / Scorpio | Statistic |
|---|---|---|---|---|---|---|---|---|
| 1,000,000 files | New full-repository workspace | Reused prepared base | 9,161.843713 | 42,244.367873 | 4.61× | Native Git worktree | 3 / 3 | Median |
| 1,000,000 files | First file usable, whole repository | First observation | 11,824.337860 | 339,282.351695 | 28.69× | Git clone --depth=1 | 1 / 1 | Single observation |
| 250 files | Complete diff · 1 file changed | First response | 1.559455 | 3.140003 | 2.01× | Standalone subdirectory Git | 1 / 1 | Single observation |
| 250 files | Complete diff · 1 file changed | Repeat response, cache retained | 1.094062 | 3.108676 | 2.84× | Standalone subdirectory Git | 2 / 2 | Median |
| 25,000 files | Dirty boolean · 1 file changed | Repeated measurement | 1.444303 | 9.661875 | 6.69× | Standalone subdirectory Git | 3 / 3 | Median |
| 25,000 files | Dirty boolean · 100 files changed | Repeated measurement | 1.527776 | 10.314854 | 6.75× | Standalone subdirectory Git | 3 / 3 | Median |
| 25,000 files | Dirty boolean · 1,000 files changed | Repeated measurement | 1.544596 | 10.269039 | 6.65× | Standalone subdirectory Git | 3 / 3 | Median |
| 25,000 files | Complete diff · 1 file changed | First response | 1.349701 | 143.243585 | 106.13× | Git sparse | 1 / 1 | Single observation |
| 25,000 files | Complete diff · 1 file changed | Repeat response, cache retained | 1.281264 | 143.197225 | 111.76× | Git sparse | 2 / 2 | Median |
| 25,000 files | Complete diff · 100 files changed | First response | 20.443732 | 151.332778 | 7.40× | Git sparse | 1 / 1 | Single observation |
| 25,000 files | Complete diff · 100 files changed | Repeat response, cache retained | 19.178516 | 151.245345 | 7.89× | Git sparse | 2 / 2 | Median |
| 25,000 files | Complete diff · 1,000 files changed | First response | 174.992904 | 230.305010 | 1.32× | Git sparse | 1 / 1 | Single observation |
| 25,000 files | Complete diff · 1,000 files changed | Repeat response, cache retained | 173.401326 | 232.129735 | 1.34× | Git sparse | 2 / 2 | Median |
Speedup = Git time ÷ ScorpioFS time. The rows include the same operation at different cache phases. A first response is a single observation and is not presented as a median. Six millisecond decimals keep the recorded format; they are not a stated error range. Dirty is only a boolean change check, not a full git status.
Full experiment reportFull experiment JSON95 performance comparisons CSV230 operation samples CSV485 Git commands CSV
Measurement definition, preparation cost, and what still needs work
Test environment: Alibaba Cloud ACS in Hangzhou, 1,000,000 files, 8,321,411,335 logical bytes, and 10 simulated history commits. The runner and the backend each have 4 CPU / 16 GiB; targeted read verification uses workers=8. Native Git CLI uses Git HTTP. ScorpioFS uses MST/2 against the same Mega2 backend. Each operation and cache phase uses the faster available Git baseline.
First usable: From the create request until the first target file is read and passes SHA-256, with the server already prepared. It does not mean the whole repository has downloaded. Explicit server preparation waited 258.59 seconds and is not part of the client speedup. Total background cost is not fully accounted and cannot be treated as zero. Background work can overlap, so it is not added directly to the explicit wait.
Workspaces and diff: A new workspace reuses its own prepared base. The Git timer is native git worktree add --detach, and both sides check the first file. A complete diff is ScorpioFS’s native interface, not a promise that git diff on the mounted directory is faster. Receiving the complete patch and SHA-256 are inside the timer. Forward and reverse application are checked independently afterward. First responses are n=1 each. Repeat responses with cache retained are the median of n=2 each. Workspaces and the dirty boolean are the median of n=3 each.
Performance boundary: This page stays with verified advantages. The first read of 1,000 files in the 25k scope took 151.68 seconds. Appending 900 files, with flush and per-file fsync, took 701.48 seconds. Updating a commit and reading 250 changed files took 61.26 seconds on ScorpioFS and 1.08 seconds on Git. Those items still need work and remain in the full report. full report
The result can be checked: 230/230 operation samples passed, 54 complete patches were verified, and 485 Git commands were recorded separately. There was no whole-repository directory walk, and no public performance number from another product is reused here.
- 230 / 230 samples passed
- 54 patches verified
- 13 verified advantages
BENCHMARK / acs-export-lifetime-1m-20261021REPORT SHA-256 / c72f3d9cf19489b2ef2394a9909070d051bc2008c9256274382a4e8442d50ff0SCORPIOFS V3 · VERIFIED 2026-10-21
Let the next agent start from the code.
From a Mega2 backend to a ScorpioFS workspace. Give your monorepo another way in.
Rust · FUSE · Monorepo · GitMono