Run the redaction engine yourself.
One endpoint: send text or a file, get deterministic placeholders back. Keep the original-to-placeholder mapping on your side, restore values in your own context, and never hand the engine anything it has to keep. Self-hosted, stateless, nothing retained.
This is the primitive, not the product. If you are guarding a live agent — tool results, prompts, replies — start atthe agent egress guard instead; it is the same engine wired into the paths an agent actually uses. If you are redacting files a human hands over, that isdocument redaction.
curl http://localhost:8090/anonymize \
-H "Content-Type: application/json" \
-d '{"text": "Account holder Jane Doe, contact jane@example.com, deploy key ghp_AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA", "profile": "egress"}'{
"text": "Account holder <PERSON>, contact <EMAIL_ADDRESS>, deploy key <GITHUB_TOKEN>",
"spans": null
}What it catches
Standard PII
Names, emails, phone numbers, SSNs, addresses, dates of birth, credit cards. The baseline every tool claims.
Identity artifacts
Bearer tokens, JWTs, OAuth codes, API keys, session tokens. The things an identity researcher knows leak.
Secrets
Private key blocks, AWS access keys, vendor key formats, and generic high-entropy patterns, with context-word scoring around each.
How a caller uses it
1. Send
POST text or a file. You get back masked text with deterministic placeholders, plus span metadata if you ask for it.
2. Keep the mapping
The engine returns offsets, never values. Your side holds the original-to-placeholder mapping: client-side, in your own store, or sealed to a vault you control.
3. Restore where it belongs
Put real values back only in the context that owns them — never on a path bound for a third party.
Profiles
One engine, different mixes of what it looks for. Pick per request.
secrets
Credentials only — vendor keys, tokens, private key blocks. For pipelines where PII is someone else's control.
egress
The standard mix: secrets plus PII. What the agent egress guard runs by default.
egress-strict
Adds network identifiers such as IP addresses, for telemetry and log paths.
FAQ
Is this a self-hosted or managed service?
Self-hosted on your infrastructure (or a managed deployment on request). Nothing retained either way: text is processed in memory and destroyed. Nothing is stored, logged, or used for training.
Does it work with streaming responses?
Not yet — the engine processes whole requests (CPU-bound, roughly 50-70 KB/s). Streaming is on the roadmap.
Why is there no accuracy number on this page?
Because a number typed into a page drifts from the engine. Our scored runs are tracked with dates, commits, and truth-set fingerprints internally, and the research behind the detectors is published at cctbp.com. Every detection claim on this site is probed against the running engine before it ships; a headline figure is not a substitute for that.
What about pricing?
No pricing announced yet. Early access users set the roadmap and lock in founding terms.