Calabi · Sanitizer
Start free trialTry free →

Calabi Sanitizer · v1

Your AI content. Metadata stripped clean.

Remove every AI provenance signal platforms scan for — C2PA credentials, generator tags, encoder fingerprints, perceptual hashes — and replace them with authentic phone-capture metadata. Every clean ships with a forensic proof report.

Start free trial → See what we strip ↓

10 free cleans · no card required · iPhone / Android / Pixel spoofing

New · Model Context Protocol

Clean AI media from any LLM.

Calabi runs a hosted MCP server — connect your AI editor and Calabi becomes a tool the model can call: scan a file, strip the fingerprints, inject an authentic phone identity, and verify the proof — all from a prompt, without leaving your editor.

Claude Code Cursor Claude Desktop Codex + any MCP client
Connect your LLM →
01

What we strip

7 detection vectors removed
Content Credentials aka C2PA · the AI signature
A cryptographically signed block of data that Sora, Midjourney, Adobe Firefly, and others attach to every output. Provably says "this was made by AI." Visible to anyone who runs a verification check. We strip the entire container.
Platforms (and journalists, fact-checkers, lawyers) look for these first. A clean file has zero of them.
Invisible watermarks the hard case
Some AI tools embed patterns directly into the pixels — invisible to you, readable by a detector. Our re-encode pass disrupts many of these patterns, but the most robust ones can survive. No tool removes every invisible watermark, and we won't pretend otherwise — the signals we fully strip are the metadata, generator tags, and encoder fingerprints.
Visible watermarks crop out. Invisible ones are the hardest signal — results vary by model.
AI-generation tags XMP, EXIF, QuickTime metadata
Hidden text fields that explicitly say "trainedAlgorithmicMedia = True" or list the model name that made the file. Many tools write these by default. We wipe every metadata field that could carry a generation tag.
A single tag like "DigitalSourceType: AI" is enough to flag your upload instantly.
Editing fingerprints encoder signatures in the bitstream
Every video encoder leaves a signature inside the file — the equivalent of a watermark from your editing software. We strip the editor's signature and replace it with the one a real phone leaves.
Detection algorithms look for "this wasn't encoded by an iPhone" as a strong AI signal.
Perceptual hash pHash · reverse-image-search defeat
Platforms compute a "fingerprint of the image content" — a small number that's the same for two visually identical files. We apply imperceptible perspective + scale + crop shifts so the fingerprint changes, defeating reverse-image search and known-content matching.
Same trick used by stock-image piracy detection. We turn it off.
Audio fingerprint tempo + sample-rate scramble
Audio gets the same treatment. Tools like Shazam and YouTube's Content ID detect audio by its waveform shape. We apply a subtle tempo shift (~0.5–1.2%) and sample-rate microshift, inaudible to humans but enough to break automated audio matching.
If your AI-generated voice or music is a known sample, audio fingerprinting catches it.
Source traces creation date, processing chain, GPS
Files generated by AI tools carry no real capture timestamp, no GPS, often no original device info. That absence is itself a signal. We strip whatever traces exist, then inject realistic ones in the next step.
"Created today, no GPS, no device" is the metadata profile of an AI-generated file.
02

What we inject

authentic phone-capture identity
Real device profile Make · Model · Software version
From a curated database of real flagship phones — iPhone 15 / 16 Pro, Pixel 8 Pro, Samsung Galaxy S24 Ultra — we inject the exact Make/Model/Software values those devices write. You can pick a specific phone or let the system rotate randomly.
A file claiming to be from a real phone passes the first metadata check.
Authentic GPS coordinates lat/lng/altitude · plausible location
We inject GPS coordinates that resolve to a real, plausible location — your choice from presets (Brickell, South Beach, Manhattan, Beverly Hills…) or paste your own lat/lng for any spot on earth. Altitude is calculated to match the local terrain.
Phones always write GPS. A file without coordinates is suspicious by default.
Realistic capture timestamp today · 1 week · 1 month · custom date
You control how old the file should look. "Today" for a fresh-feeling post, "1 month ago" for a video from your trip last month, or a custom date for any narrative. We inject the timestamp into every date field a phone would touch (CreateDate, ModifyDate, CreationDate).
Inconsistent dates across fields = obvious tampering. We keep them all consistent.
Phone encoder signature H.264 · phone-style compressor name
Where the original file was encoded by editing software, the output now reports the encoder a phone uses. The compressor name in the container is updated, and the editor's version string is suppressed from the bitstream.
The encoder name is one of the cheapest signals to scan for. We make sure it matches a phone.
Timezone + local format Apple ISO · UTC offset · per-device format
Apple devices format dates differently than Android. Both use timezone offsets. We match the format to the device profile so an "iPhone 16 Pro" file uses Apple's ISO format and an Android file uses the Android style.
Format mismatch ("Android file written by Apple") fails sanity checks.
03

The three stages

strip → inject → verify
i. Strip All 7 detection vectors removed. Pixels re-encoded with controlled noise so invisible watermarks are broken.
→
ii. Inject Authentic phone capture identity — device, GPS, timestamp, encoder — written across every metadata field.
→
iii. Verify The same forensic scan platforms use is run against the output. You see the result before you download.
04

Honest answers

FAQ
Does it work on every AI tool?
Yes — the pipeline is tool-agnostic. We don't try to detect what tool made the file; we strip every signal that any AI tool could have left. Tested against Sora, Veo, Kling, Runway, Seedance, Midjourney, Stable Diffusion, DALL-E, Adobe Firefly, Google Gemini's image gen, Nano Banana, and more.
Will the output still look like the original?
Yes. The visible content is preserved. The changes are at the metadata level (invisible) and at the pixel level (imperceptibly small noise + perspective shifts that defeat fingerprinting but never affect what you actually see).
What does "Aggressive" or "Maximum" strip strength do differently?
Standard handles the most common detection paths and is what most users need. Aggressive adds a perceptual-hash defeat layer (subtle perspective + scale shift) and a stronger audio scramble — use when you're worried about reverse-search matching. Maximum adds a color-grade microshift on top — useful for highly recognizable scenes where every signal needs to be perturbed. Maximum takes slightly longer to process.
Why does the Verify tab show a green check?
The Verify panel runs the same metadata scan platforms use, against the output file. Green = no AI signals detected, encoder fingerprint matches a real phone, identity injected successfully. If anything leaked, the panel shows it. You see the proof before you download.
Can platforms detect the file was "cleaned"?
The output presents as a normal phone-recorded file — that's the whole point. There's no tell-tale "this was sanitized" marker because we don't add one. Every change we make is something a phone could plausibly have produced naturally.
What file types?
Videos: .mp4, .mov, .m4v, .webm. Images: .jpg, .png, .webp, .heic, .heif. Up to 200 MB per file. Batch upload supported.
What happens to my uploads?
Files are processed and held for 1 hour, then auto-deleted. We never train on user content, never read it, never share it. The pipeline is fully automated end-to-end. Output files are served over a TLS-encrypted Cloudflare tunnel.
Custom GPS — can I pick any location?
Yes. Pick "Custom" in the location dropdown and paste lat/lng for any spot on earth. Add a place name (e.g. "Lake Como, Italy") for the identity string. Altitude is auto-set.
05

Simple pricing

free to start · pay when you scale
Free trial
$0
  • 10 cleans total
  • Standard strip strength
  • All output formats
  • Forensic proof on every clean
  • No card required
Start free trial →
Weekly
$12/week
  • Unlimited cleans
  • All strip strengths (Standard / Aggressive / Maximum)
  • Quality presets (IG / TikTok / YT)
  • Custom GPS + capture dates
  • Cancel any time
Choose Weekly →
Save 27%
Monthly
$34.99/mo
  • Unlimited cleans
  • Priority processing queue
  • Everything in Weekly
  • Batch up to 50 files at once
  • Saves $13/month vs weekly
Choose Monthly →
Drop a file. See the forensic report before you download.
Start free trial →