Blackbar — local-first redaction for confidential documents

The redaction

Removal,
not covering.

A black rectangle drawn on top is not a redaction — the words are still in the file, one copy-paste away. Blackbar deletes the underlying text from the document, destroys image pixels under every box, and scrubs metadata and hidden carriers (field codes, comments, tracked changes).

Then it checks its own work: every export is re-read by independent extractors. If anything you accepted can still be recovered, the export is refused.

settlement_agreement.pdf ✓ verified safe

Before

This Agreement is between Marcus Webb of Acme Holdings LLC for the sum of $2,450,000.00, dated March 14, 2025.

After

This Agreement is between Marcus Webb of Acme Holdings LLC for the sum of $2,450,000.00, dated March 14, 2025.

Text deleted from the content stream · copy-paste recovers nothing.

The app

Upload. Detect. Review. Verify.

A local web app you run on your own machine. Drop a document; detection proposes every sensitive span, grouped by category. Accept, reject, draw your own boxes — then export, gated by the safety check.

127.0.0.1:8000 · blackbar (local)

Blackbar review workspace: proposals grouped by category on the left, the document with live black redaction bars in the middle, and the server-rendered redacted preview on the right.

Nothing redacts until you say so

Every detection is a proposal. Accept or reject each one, resize boxes, or redact every occurrence of any term with one click.

See exactly what ships

A server-rendered preview shows the real redacted output beside the original — what you review is byte-for-byte what you export.

A manifest you keep

Each export writes a JSON manifest: input/output hashes, every proposal and your decision, and the verification verdict.

Built for confidential work

Quietly thorough
where it counts.

Every document you handle

PDF, Word, PowerPoint, Excel, HTML, Markdown, and images. Scanned pages and images embedded inside documents are OCR'd locally and redacted too.

Found in seconds, no AI

Names, parties, addresses, emails, phones, amounts, dates, SSNs, cards, IBANs, matter IDs — checksum-validated where it counts. A long contract scans in seconds.

True visual redaction

Content is removed from the file and image pixels destroyed under each box — never a rectangle drawn over recoverable text.

Verified, not promised

Independent extractors (and a re-OCR pass for scans) confirm accepted text is gone. If it isn't, the export is refused and destroyed.

Readable replacements

Optionally swap each entity for a stable, realistic fake — entity-linked so the same original maps to the same fake everywhere and the document still reads for downstream AI.

Nothing leaves your machine

No telemetry, no accounts, no cloud. The optional deep scan runs locally via Ollama. Non-English names (Korean · Japanese · accented Latin) detected too.

Engine

Deterministic by default.
Optional local AI.

Detection is regex recognizers + contract-structure extraction + spaCy NER + your watchlist — fast, repeatable, and entirely offline. For messy or unusual documents, one button adds a local LLM via Ollama. Either way, it all runs on your machine.

Regex · Structure · NER Default seconds · no AI required · no network · + optional Ollama deep scan

Two ways to run it

Careful review,
or one command.

The review app is the intended workflow for real redactions. For batch or scripted runs, the same pipeline — and the same safety gate — runs headless.

	Review app	Headless CLI
How you run it	`redactor serve` → localhost	`redactor redact file`
Human review	Accept / reject each span	Accepts all detections
Best for	Careful one-off redactions	Batch & scripted runs
Formats	All	All
Extraction-safety gate	Always	Always
Runs locally	Yes	Yes

Get started

Clone, set up, run.

macOS or Linux · Python 3.10+ · Node 18+ · LibreOffice for Office/HTML/Markdown proofs · Ollama and tesseract optional.

Get the code and set it up once.

git clone https://codeberg.org/russkysong/blackbar.git
cd blackbar
python3 -m venv .venv
.venv/bin/pip install -e '.[dev]'
.venv/bin/python -m spacy download en_core_web_md
cd frontend && npm install && npm run build && cd ..

Run the review app and open it in your browser.

.venv/bin/redactor serve
# then open http://127.0.0.1:8000

Or redact headless, one command — same safety gate.

.venv/bin/redactor redact contract.pdf                 # fast, no AI
.venv/bin/redactor redact contract.docx --replace    # realistic pseudonyms

Get it on Codeberg →

Removal,not covering.