CaptureMann

Status: 0.2.5 (macOS 13+, Apple Silicon and Intel). Windows, then Linux follow. Most results are tested with scripted fakes, plus the maintainer's own use on one Mac; there are few stored hardware reports. Speaker labels are measured on public English conversation clips only (diarization error 17 % on clips held back, report); Vietnamese is not measured yet. Intel: the build installs, opens and asks for permissions on a real Intel Mac, but starting a recording crashed one (issue 16); 0.2.3 and later have a fix that is not confirmed yet. What is still unverified is tracked in specs/04-delivery-plan.md under "Verification debt" and in docs/project/status.md.

Features

• Screenshots of a display, a window or a region, with an annotation editor and Auto-redact for personal data and secrets (see below).
• Screen recording up to the display's native resolution (4K/5K) at 30 or 60 fps, with a webcam bubble, microphone (noise suppression), system audio and cursor telemetry. Pause, resume, and crash recovery that loses at most two seconds.
• Editor: trim, cut, split, speed, backgrounds, zoom regions, restylable cursor, 9:16 and 1:1 reframe. Export to MP4, platform presets, GIF or WebM.
• Audio recording: meetings, interviews, lectures and voice notes from the microphone and, optionally, system audio. A compact recording bar with a level meter, markers (press M), a consent reminder, and a WAV master that survives a crash. An audio editor with a waveform, a transcript, trim and cut, and export to M4A or WAV with loudness levelling.
• Local AI, all on your Mac: Whisper transcripts and captions (SRT, VTT, TXT, or burned in) with Large v3 Turbo as the recommended model, a caption editor, translation to English with Whisper and to other languages with the built-in text model, and zoom and silence-cut suggestions. Models download only when you press Download.
• Speaker labels: who spoke when, from the voices, with rename and merge and a progress bar on long recordings. The two small models ship inside the app; nothing is sent anywhere. Tell the app which languages you speak (for example vi, en) so a short recording is not heard as a related language.
• Meeting minutes and reports: summary, minutes, one-on-one notes, interview notes, product recap and lecture notes written by a built-in text model (Gemma 4 E2B by default). Every item links to the moment it was said; export as Markdown or text. The model runs through a bundled llama-server; no cloud and no account. The picture below shows a report with sample data.
• Library of self-contained capture bundles, stored in a folder you choose.
• Share: save to local folders or Google Drive, publish to YouTube (private by default).
• For agents: 74 MCP tools plus resources and prompts, and an identical capturemann CLI. Risky actions (start recording, publish, delete) need an on-screen confirmation.
• Private by design: no telemetry, and no network traffic unless you connect a destination or download a model. You bring your own Google credentials; secrets live in the macOS Keychain.
• English and Vietnamese interface.

Find personal data before you share a screenshot

Auto-redact reads the text in a screenshot with the recognizer built into your Mac (Apple Vision) and proposes blur rectangles over the personal data and secrets it finds. You review the proposals (masked previews, dashed boxes) and accept them one by one or all at once; nothing is applied automatically. It runs on your computer and uses no network. The same function is available to agents (MCP suggest_redactions, which only proposes) and to the command line (capturemann annotate <id> --suggest-redactions). The user guide has the details: Auto-redact.

It finds formats and checksums, not meaning: a payment card number (Luhn check and issuer prefix), a Social Security number in an issuable range, a routing number with a valid checksum, a private-key block, a PASSWORD=... line. It does not find names, addresses or dates. The kinds, generated from the rule data (a test fails when this table and the code differ):

Kind	Group	What it finds	Standards that name it	On by default	Check used
email	Contact	Email addresses ([email protected]).	HIPAA 6	yes	Shape only
token	Credentials and secrets	API keys, JWTs and other long random-looking strings: known provider shapes, or a mix of 24 or more letters and digits (lower confidence).	GLBA / state breach laws (credentials)	yes	Known prefix and length, or letter-and-digit mix
private_key	Credentials and secrets	PEM private-key blocks, every line from the BEGIN line to the END line.	GLBA / state breach laws (credentials)	yes	BEGIN and END lines
secret_assignment	Credentials and secrets	The value in NAME=value or "name": "value" pairs whose name looks secret (password, token, key and similar).	GLBA / state breach laws (credentials)	yes	Secret-looking name, value is not a placeholder
url_secret	Credentials and secrets	A password inside a URL, or a secret value in a URL query string.	HIPAA 14; GLBA / state breach laws (credentials)	yes	Userinfo password or secret-looking parameter name
card	Payment cards	Payment card numbers of 13 to 19 digits; masked numbers are skipped.	PCI DSS; HIPAA 10; GLBA	yes	Luhn + issuer prefix + length
card_extra	Payment cards	An expiry date or security code beside a detected card number (found by position, no label needed).	PCI DSS (sensitive authentication data)	yes	Valid month, or 3 or 4 digits, beside a detected card
ssn	Identity and tax numbers	US Social Security numbers written NNN-NN-NNNN or with spaces; masked forms and bare nine-digit numbers are skipped.	NIST SP 800-122; HIPAA 7; GLBA	yes	Issuable ranges (no 000, 666 or 9xx area, no 00 group, no 0000 serial)
itin

The "Standards that name it" column maps each kind to the public lists that mention that kind of data (NIST SP 800-122, HIPAA Safe Harbor identifiers, PCI DSS, GLBA and state breach-notification laws). This is a mapping of kinds, not a compliance claim: CaptureMann is not a compliance tool, and using it does not make a screenshot compliant with any of them.

How well it works (measured, synthetic data)

All numbers below come from a stored report and are measured on synthetic screenshots made by this project (300 images, fixed seed 20261030, 13 screen layouts, light and dark, 1x and 2x, some degraded): 211 images from 9 layouts (dev, where the rules were tuned) and 89 images from 4 other layouts (held_out, read once after everything was fixed). The images contain made-up data whose formats the rules know, so these figures are an upper bound for real screenshots, not a measure of real-world accuracy. Recognizer: Apple Vision, on an Apple M4 Pro, macOS 26.6.2. Date: 2026-10-30.

Apple Vision, synthetic	dev: 0.2.3 (4 kinds)	dev: 0.2.5	held-out: 0.2.3	held-out: 0.2.5
Items scored	1,367	2,275	436	1,410
Found (any kind)	92.2 %	95.3 %	94.5 %	98.6 %
Precision (proposals that were real items)	68.8 %	92.0 %	69.2 %	94.4 %
False alarms per image	2.725	0.417	2.056	0.831
False alarms per image with nothing to find	2.000	0.746	2.074	0.889
Leak rate (still readable after the blur)	4.2 % (58 of 1,367)	2.1 % (47 of 2,275)	3.4 % (15 of 436)	0.5 % (7 of 1,410)

The old and new columns score different sets of items (the old detector knew 4 of the 20 default kinds), so the "found" rows are not like for like. With perfect text instead of Apple Vision (this separates rule errors from recognizer errors) the 0.2.5 rules find 99.2 % at 98.5 % precision on dev and 100 % at 95.9 % on held_out. Report: 0.2.5 final run (it includes the 0.2.3 detector on the same held-out layouts), with the earlier 0.2.3 baseline and interim 0.2.5 run; per-kind, per-layout, per-font-size and per-language tables and the lookalikes that raised false alarms are in them.

What the numbers say plainly. The false-alarm target is not met: the goal was below 0.2 false alarms per image with nothing to find and 0.2.5 shows 0.746 (dev) and 0.889 (held_out). The generic long-token rule (unchanged from 0.2.3 and kept on) is the main cause, for example by proposing full git hashes. npi is off by default because only 30 % of its proposals were real. Private-key blocks are the weakest default kind with Apple Vision (65.5 % of key lines covered, 15.5 % still readable after the blur, dev only: the held-out layouts contain no keys or secret URLs). Nothing here has been checked on real screenshots yet (that is the maintainer's manual cross-check, tests/manual/RED-1.md).

To reproduce on your own Mac (builds the synthetic set from the seed, no network; run uses the system recognizer):

cargo run -p cm-render --example redaction_bench --release --locked -- gen --out /tmp/redaction-set
cargo run -p cm-render --example redaction_bench --release --locked -- run --set /tmp/redaction-set --detector-only --split dev
cargo run -p cm-render --example redaction_bench --release --locked -- run --set /tmp/redaction-set --recognizer system --split dev

What it does not do

• It finds formats and checksums only. Names, street addresses and dates are not found, and there are no label-based kinds: a CVV or an expiry date with no card number on the screen is not proposed, and a bare value under a "Password" or "SSN" label is not found by the label (a NAME=value pair with a secret-looking name is).
• Text the recognizer cannot read stays. Small, low-contrast or unusual text can be missed, which is what the leak rate measures. Private-key blocks are the weakest default kind with Apple Vision.
• Bare ten-digit Vietnamese mobile numbers (0934851497) are not found, and passwords made only of letters (letmein) are not proposed.
• Some lookalikes are proposed anyway (in the dev split: hash strings, timestamps, ten-digit ids, tracking numbers); that is why you review.
• There is no Windows or Linux recognizer yet; the Auto-redact button is hidden there.
• It is not a compliance tool. A blurred screenshot is not proof that nothing sensitive remains: look at the result before you share it.

Install

Requires a Mac with macOS 13 or later, Apple Silicon or Intel (Apple menu > About This Mac shows the chip). There is one download per chip. The Intel build (0.2.1 and later) is built and checked on Apple Silicon. It installs and opens on a real Intel Mac, but starting a screen recording crashed a MacBook Pro 16" 2019 in 0.2.2 (issue 16); 0.2.3 catches the error and retries with simpler settings, which is not confirmed on that Mac yet. The speech and text models may be slow on older Intel Macs. Please report problems. The app is about 180 MB (Apple Silicon) or 190 MB (Intel). The AI models are not in it: they download only when you press Download (about 4.4 GB for the recommended speech model, Large v3 Turbo, plus the recommended text model, Gemma 4 E2B). The two small speaker models are inside the app.

With Homebrew:

brew install --cask craft-mann/tap/capturemann

Homebrew picks the right build for your chip. This also links the capturemann command line into Homebrew's bin. Update with brew upgrade --cask capturemann.

From the release page:

1. Download CaptureMann-v<version>-macos-arm64.dmg (Apple Silicon) or CaptureMann-v<version>-macos-x86_64.dmg (Intel) and its .sha256 file from Releases.
2. Check the download (in the folder with both files): shasum -a 256 -c <the DMG's name>.sha256 must print OK.
3. Open the DMG and drag CaptureMann to Applications.

The release DMG is signed with a Developer ID certificate and notarized by Apple, so it opens normally. A build you make yourself from source is signed locally or ad hoc and is not notarized: Gatekeeper blocks it at first launch, so right-click the app and choose Open, or run xattr -dr com.apple.quarantine /Applications/CaptureMann.app. Updates are manual: download the new DMG. The capturemann command line is inside the app; to use it as capturemann, link it once: sudo ln -s /Applications/CaptureMann.app/Contents/MacOS/capturemann /usr/local/bin/capturemann. To build it yourself instead, see building from source. More on verifying a download: docs/development/releasing.md.

Quick start

make run-bundle                 # builds, signs and opens target/bundle/CaptureMann.app
# the CLI and MCP server are inside the bundle, in Contents/MacOS
target/bundle/CaptureMann.app/Contents/MacOS/capturemann capture screenshot

Connect an AI agent over MCP (stdio):

{ "mcpServers": { "capturemann": { "command": "/path/to/CaptureMann.app/Contents/MacOS/capturemann", "args": ["mcp"] } } }

See specs/02-agent-interface.md for the tools, permission tiers and the CLI.

Build from source

Rust 1.95, Xcode command line tools and cmake:

cargo test --workspace --locked
make bundle

Full instructions, the bundled LGPL FFmpeg and macOS permissions: docs/.

Documentation

• docs/: building, permissions, plan, design, spikes
• specs/: product spec, technical design, agent interface, integrations, features
• CONTRIBUTING.md, SECURITY.md, CODE_OF_CONDUCT.md, CHANGELOG.md

Contributing

Contributions are welcome, especially new storage providers, publishing platforms and translations. Read CONTRIBUTING.md first. Security problems: please report them privately, see SECURITY.md.

Links

CaptureMann is a CraftMann app. Contact · Privacy policy · Source

License

Apache-2.0, see LICENSE. Third-party notices: NOTICE and THIRD_PARTY_LICENSES.md. The bundled FFmpeg is an unmodified LGPL build.

The Apache-2.0 licence covers the code. It does not grant rights to the CaptureMann or CraftMann names or logos (design/brand); forks should use their own name and marks.