cloakroom 0.4.2

Get started · Blocked prompt? · Results · How it works · Dashboard · FAQ · Known limits

Cloakroom swaps your secrets and personal data for placeholders before Claude sees them, and puts the real values back in the files Claude writes.

The only Claude Code privacy tool that catches personal data, masks instead of blocking, and learns from your corrections.

Runs on your machine, sets itself up on first start, works on Mac, Windows and Linux. Free and open source. Young: expect rough edges.

Hand Claude the files you'd never paste into a chat:

• Bank statements (PDF, scan or CSV): "sort a year of spending by category." Claude sees amounts and merchants, not your name, address or account number.
• Support tickets: "draft replies to this week's tickets." Claude writes to a stand-in name (PII_PERSON_1); the drafts saved on your disk have the real names.
• Medical or HR records: names, birth dates, addresses and ID numbers are masked, so Claude can summarise or tidy a record without knowing whose it is.
• .env files and logs: Claude fixes the config without ever seeing the password or API key.
• Scans and photos of documents: read on your machine (no upload) and masked like any other file.
• Learns from you: mark a false alarm once in the dashboard and a rule stops it next time.

_{Claude writes the reply. The customer's name never left your machine.}

Get started

Inside Claude Code:

/plugin marketplace add shanjeevrajendran/cloakroom
/plugin install cloakroom@cloakroom

That's it. Restart Claude Code and Cloakroom sets itself up: keys, card numbers and ID numbers are covered within seconds; names and addresses a few minutes later, once the detection model has downloaded (about 1 GB on a Mac, up to about 4 GB elsewhere). It needs uv, pipx or Python 3.10 or newer. Tested from clean installs on macOS, Windows and Ubuntu. Progress: ~/.cloakroom/bootstrap.log; details in Setup.

Then, in a terminal:

cloakroom doctor       # checks Python, the key store, the model and the service
cloakroom dashboard    # opens your alerts in the browser

Prefer to install by hand: uv tool install "cloakroom[model,torch,ocr]" (Python 3.10+; the part in brackets adds the detection model and OCR), then add the plugin as above.

How it works

  Claude Code reads a file or runs a tool
        │   "Nadia Petrov called about…"    sk_live_51Hx9Q…
        ▼
  ┌────────────────────────────────────────────┐
  │ Cloakroom  (on your machine)               │
  │   rules + local model  ->  find the values │
  │   encrypted vault      ->  remember them   │
  └────────────────────────────────────────────┘
        │   "PII_PERSON_1 called about…"    PII_SECRET_1
        ▼
  the model works with placeholders only
        │   Write reply.md: "Hi PII_PERSON_1,"
        ▼
  Cloakroom puts the real value back, locally
        │
        ▼
  reply.md on your disk: "Hi Nadia Petrov,"

Details: docs-HOW-IT-WORKS.md.

Dashboard

Where your labels become rules: false alarms fall as you label them, and the detection model itself never changes.

cloakroom dashboard opens a page served by Cloakroom itself on 127.0.0.1, signed in with a one-time link.

• Alerts: every masked, withheld or blocked call, with the kinds found and how long the check took. Open one to see its placeholders; the real values show only when you click Reveal.
• Your labels: mark an alert correct, a false alarm, or type what it missed.
• Rules from your labels: a false alarm suggests "never mask this", a miss suggests "always mask this". You approve each one, or later let rules apply by themselves after a replay test against your past labels; rules that would unmask a key, password, card or ID number always wait for you. Undo puts everything back. The detection model never changes. Details: the handbook.
• Private by design: the activity log holds no values or text; labels and rules are encrypted like the vault.

When Cloakroom blocks your prompt

Claude Code lets Cloakroom block a prompt but not rewrite it, so a prompt holding a key, an email, a phone, a card or an ID number is stopped. Names alone and plain dates go through. The block message says what it found, and how to get past it:

Files and command output stay masked the whole time. Only you can do this: Claude never sees these commands and can't change Cloakroom's settings. Usually better still: put the detail in a file and ask Claude to read it. It gets masked on the way in and the real values come back when Claude writes.

Why Cloakroom

Claude Code reads whatever your task touches: .env files, logs, customer exports, support tickets. All of it goes to the model. Secret scanners catch keys with a known shape, but a customer's name in a ticket or an address in a log looks like ordinary text to them.

• Catches personal data, not just keys. A small detection model on your machine, plus rules, finds names, emails, phone numbers and addresses that have no fixed shape. On fresh test sets it let through 0 of 1,005 of them; the other tools we tested let through 119 to 992 (Results).
• Claude keeps working. Placeholders stay the same for the whole session, and the real values are put back when Claude writes a file or runs a command locally.
• Runs on your machine. Detection, OCR, the encrypted vault (where the real values wait) and the put-back are all local. The only network call is the one-time model download (the OCR models ship inside the package).
• Mac, Windows and Linux. One install on all three, with the vault key in each system's own secure store: Keychain on macOS; Credential Manager on Windows (runs in PowerShell or Git Bash, tested on Windows Server); Secret Service on Linux (tested on Ubuntu 24.04, plus automatic tests on every change).
• No graphics card needed. Each file or command is checked in a fraction of a second on an ordinary laptop (cloakroom doctor prints your numbers).
• Catches encoded output. If a command prints data in encoded form (base64, hex dumps from xxd or od), Cloakroom decodes it and checks it too.
• If it can't check, it doesn't let things through. When Cloakroom isn't ready, Claude's file reads and commands wait instead of passing unchecked. Your prompts still get a quick rules check, with a warning, so you're never locked out.

Results

• Good at: personal data (names, emails, phone numbers, addresses) and secrets in odd places. On fresh test sets it let through 0 of 1,005 personal details; the other Claude Code tools we tested let through 119 to 992. It masks, so Claude keeps working; one other tool blocks the whole step instead.
• Weaker at: False alarms. Cloakroom is tuned to mask rather than risk a leak, so it masks more harmless text than secret-only scanners; Claude still gets the job done with a placeholder. Label a false alarm in the dashboard and a rule stops it next time.
• What a false alarm costs: Claude sees a placeholder where it needed the real text. Nothing leaks.

Every number comes from a set run once before any tuning on it unless it says otherwise. All tools at default settings, offline, tested 12 to 14 October 2026. maisecrets' maintainers reproduced our numbers and pointed out where the older sets were unfair to them (keys in formats vendors don't issue, and the regions setting); the key-format set and the setting below are the fix. The maintainers of the compared tools were contacted before publishing. Method, versions and every run: bench/RESULTS.md.

Made-up agent traffic (five sets written by a different model: chats, logs, config files, spreadsheets, support tickets): 1,339 key and personal values (the rest are in RESULTS.md) and 200 safe look-alikes. Fewer is better. These sets have been seen by now, so Cloakroom's second row is not an honest score; the first row is its run before any tuning on them.

Some of these tools only aim at secrets, so their personal-data column shows scope as much as quality. Cloakroom's one leak today: a date of birth in a dating-profile export.

Keys in the formats vendors really issue (a new set written after the first comparison: 15 key types, passwords with no fixed shape, encoded emails; 77 values, 15 look-alikes):

Real passwords and keys from public code (Samsung CredData): 14,471 real secrets and 15,422 harmless look-alikes, each line checked alone, the way a tool output reaches Claude. One repository is almost all cipher test vectors (sample keys for testing encryption code), which maisecrets' keyword rule is built to catch: first number without it, in brackets the whole set.

Without the test-vector repository, Cloakroom catches the most; with it, maisecrets does, and maisecrets is the most precise either way (of what it masks, 86 to 91 in 100 are real secrets, against Cloakroom's 71). Cloakroom masks far more harmless look-alikes (3,903 against maisecrets' 997 on the whole set): false alarms are Cloakroom's main open problem (precision, recall and F1 for every tool).

On public datasets we didn't write (run once; details in bench/public/RESULTS.md):

On Gretel, over half of those extra masks are number strings Cloakroom treats as account or reference numbers; the rest are mostly names and addresses Gretel leaves unlabelled.

Ways to run

• Full mode (default): rules plus the local model. Catches keys and personal data written as plain text.
• Rules mode (CLOAKROOM_MODE=rules): no model, instant, installs with plain uv tool install cloakroom. Catches keys, card numbers and IDs, but misses most names and addresses.
• Outside Claude Code: cloakroom check FILE shows what would be masked, cloakroom explain FILE shows why, cloakroom mask FILE prints what Claude would see.
• Dry run (CLOAKROOM_DRY_RUN=1): lets everything through and logs what it would have done.
• Settings in ~/.cloakroom/config.json: "allow" for your own exceptions, "mask_all_dates": true to mask every date (by default only birth dates and dates in someone's record), "mask_own_username": true to mask your own login name in paths too, "allow_images": true to let pictures through unchecked (with a warning).

Exceptions, backends, CI and troubleshooting: the handbook.

FAQ

Does it send my data anywhere? No. It downloads the detection model once from Hugging Face (a pinned version); the OCR models come inside the package. Everything else runs on your own machine (127.0.0.1).

How much slower is Claude? Each tool call is checked by a small local service: about 20 to 70 ms on Apple Silicon, where the model runs on the built-in GPU (an M5 Max to an 8 GB M1 Air), 60 ms on an Apple M-series CPU, about 300 ms on a 2-core cloud VM. Without an Apple Silicon or NVIDIA GPU, a fast CPU version of the model is built on first start. The first run downloads the model once (about 1 GB). cloakroom doctor prints your numbers.

Will it mask things it shouldn't? Sometimes, see above. Documented examples (Stripe test cards, example.com, fictional 555 numbers), plain dates and job titles are left alone. Mark a false alarm on the dashboard and it suggests a rule, or add exceptions to allow in ~/.cloakroom/config.json. cloakroom check FILE previews masks.

What doesn't it protect? It stops accidental exposure; it isn't a sandbox. Out of scope: a prompt injection (text in a file that steers Claude) set on smuggling data out piece by piece, what you type into a prompt (a prompt can be blocked but not rewritten, and names alone go through), and Claude Code's own local logs. Details: docs-LIMITS.md.

What about PDFs, scans and photos? Cloakroom pulls the text out on your machine and masks it like any file; Claude gets the text, never the PDF or the picture. Scanned pages and photos (png, jpg) are read by OCR (reading text from an image), running on your CPU with its models inside the package: no download, no cloud, no account. A scan it can't read in time, encrypted PDFs, figures in a notebook and other pictures are held back, and Claude is told so; to show Claude a picture anyway, attach it to your own prompt. Page and time limits: known limits.

How do I remove it? See Uninstall.

More

• Write-up: why I built it and how masking, OCR, restore and the dashboard fit together
• How it works: the detectors, the known-safe filter, the vault and the restore
• Changelog · Setup details · Known limits · Benchmark
• Found a way around it? Please report it privately: SECURITY.md

Author

Made by Shanjeev Rajendran · GitHub · LinkedIn

Credits

Detection model: PII-Tracer. The one-string hook command that runs in both bash and PowerShell comes from maisecrets (Apache-2.0). Licensed Apache-2.0.