antiaimark

English | 简体中文 | Español | Français | Русский

Detect and strip AI provenance marks from text, images, documents, videos and audio — invisible Unicode steganography, C2PA/EXIF/XMP image metadata, container metadata (PDF, DOCX, ODT, SVG, HTML, Markdown, best-effort video) — with CLIs, an HTTP service + web UI, an MCP server for AI IDEs, and background auto-clean. Pure Go, static binaries, no runtime dependencies.

Features

• Text (Layer A) — zero-width characters, bidi overrides, tag characters, homoglyph spaces, private-use planes; byte-exact round-trip of non-UTF-8 input
• Images — PNG/JPEG/WebP metadata: C2PA/JUMBF manifests, XMP digitalSourceType=trainedAlgorithmicMedia, generator text chunks; pixel data untouched
• Containers — PDF (exiftool + qpdf when available), DOCX/ODT internals, SVG metadata blocks, HTML meta/JSON-LD, Markdown frontmatter
• Video & audio — best-effort: C2PA uuid/JUMBF box scan, QuickTime ©too atom, marker scan (Suno/ElevenLabs/MusicGen…), exiftool -all= strip
• Vendor keywords — OpenAI/Imagen/Firefly/Midjourney/Stable Diffusion/ FLUX/Ideogram/Recraft/Grok + 豆包·即梦/腾讯混元/通义万相/可灵/智谱/文心一格/海螺… (CMS tags like WordPress are preserved)
• HTTP + web UI — JSON API (/inspect /clean), drag & drop upload with one-shot download for images and videos

• MCP server — native tools inside Claude Code/Desktop, Cursor, Windsurf, Cline, Continue, Zed…
• 5 languages — en / zh / es / fr / ru across CLIs, HTTP errors, web UI and MCP descriptions
• Background auto-clean — disk-space threshold, configurable period

Quickstart

go build ./...          # build everything
go test ./...           # run the suite
./deploy.sh build       # or: build the single binary into bin/
./bin/antiaimark server # HTTP + web UI on 127.0.0.1:8765

Open http://127.0.0.1:8765/ and drag an image or video in. CLI examples:

./bin/antiaimark inspect-file photo.png --json    # unified inspect (auto-routes)
./bin/antiaimark clean-file   doc.docx            # writes doc.cleaned.docx
./bin/antiaimark clean-text   notes.txt --lang zh # localized CLI messages
./bin/antiaimark audit-dir    ~/blog              # aggregate directory audit

Deployment

./deploy.sh covers every path (also via make package, make install-systemd):

Command	What it does
./deploy.sh build	build all binaries for the host platform into bin/
./deploy.sh build-linux [amd64|arm64|386]	cross-compile static linux binaries
./deploy.sh package [arch]	self-contained tarball in dist/ (binaries + README + deploy/)
./deploy.sh docker-build [tag]	build the distroless Docker image
./deploy.sh docker-run	run the image with production defaults (loopback, read-only, tmpfs)
sudo ./deploy.sh install-systemd	Linux bare-metal: binaries + dedicated user + env file + hardened systemd unit
sudo ./deploy.sh uninstall-systemd	stop and remove the unit

Bare-metal flow on a linux server:

./deploy.sh package amd64                  # on your workstation
scp dist/antiaimark-*-linux-amd64.tar.gz server:
ssh server 'tar xzf antiaimark-*.tar.gz && cd antiaimark && sudo ./deploy.sh install-systemd'
# configure: sudoedit /etc/antiaimark.env   (port, API key, auto-clean…)
sudo systemctl restart antiaimark

Docker alternative: docker compose up -d (loopback bind, healthcheck, read-only rootfs, dropped capabilities; knobs via environment).

Configuration (env file / environment)

Variable	Default	Meaning
ANTIAIMARK_SERVER_HOST	127.0.0.1	bind address (loopback-only unless behind a proxy)
ANTIAIMARK_SERVER_PORT	8765	port
ANTIAIMARK_SERVER_API_KEY	empty	require Authorization: Bearer <key> when set
ANTIAIMARK_LANG	system locale	en zh es fr ru
ANTIAIMARK_AUTO_CLEAN	0	1 enables the background janitor
ANTIAIMARK_AUTO_CLEAN_INTERVAL	15m	check period
ANTIAIMARK_AUTO_CLEAN_THRESHOLD	11	free-space % that triggers cleanup
ANTIAIMARK_AUTO_CLEAN_TTL	24h	download retention before eviction

The janitor only ever deletes this service's own wm-* temp directories and expired downloads — nothing else on disk; directories younger than 1h are protected so in-flight requests are never disturbed.

HTTP API

Method	Path	Purpose
GET	/health /capabilities /openapi.json	liveness, tool availability, OpenAPI 3.0.3
POST	/inspect / /clean	base64 file in, findings/cleaned bytes out
GET	/	web UI (drag & drop, 5 languages)
POST	/api/upload → GET /api/download/{token}	multipart upload, one-shot cleaned download
GET	/api/i18n?lang=zh	message catalog for the UI

AI IDE integration (MCP)

claude mcp add antiaimark -- /abs/path/to/bin/antiaimark mcp
# Cursor / Windsurf / Cline mcp.json:
{ "mcpServers": { "antiaimark": { "command": "/abs/path/to/bin/antiaimark", "args": ["mcp"] } } }

Two transports are supported:

• stdio — bin/antiaimark mcp, the command form above (Claude Code, Cursor, Windsurf, Cline, Continue, Zed, …).
• Streamable HTTP — register the running HTTP service (antiaimark server) as a remote MCP server, no binary path needed:

{ "mcpServers": { "antiaimark": { "type": "http", "url": "http://127.0.0.1:8765/mcp" } } }

Tools: capabilities, inspect_file, clean_file, inspect_text, clean_text — descriptions localize to the IDE language.

Step-by-step registration for every IDE/agent (Claude Code/Desktop, Cursor, Windsurf, Cline, Continue, Zed, VS Code Copilot, JetBrains, Trae, Codex, Gemini, Amazon Q, WorkBuddy, DeepSeek Harness): see docs/MCP.md. Prebuilt binaries for all platforms are published to GitHub Releases and installed with the one-liners below:

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/n0bele/AntiAIMark/main/scripts/install.sh | bash
# Windows (PowerShell)
irm https://raw.githubusercontent.com/n0bele/AntiAIMark/main/scripts/install.ps1 | iex

Claude Code plugin marketplace

Installable as a Claude Code plugin (one command, auto-updates):

claude plugin marketplace add n0bele/AntiAIMark
claude plugin install antiaimark@antiaimark

Or in the /plugin UI: Marketplaces → add n0bele/AntiAIMark, then install antiaimark. See plugins/antiaimark/README.md for details. To get listed in the Claude Code Discover tab / community marketplace, follow docs/plugin-submission.md.

Optional ML harnesses

Pixel removal (CtrlRegen / MarkDiffusion), SynthID scoring and MarkLLM text detection run as external adapters when NOAI_WATERMARK_DIR, MARKDIFFUSION_DIR, REVERSE_SYNTHID_DIR or MARKLLM_DIR point at checkouts; without them the core works standalone and /capabilities reports them as absent.

Architecture

Core library plus thin facades (CLIs / HTTP / MCP) — see README-ARCHITECTURE.md for the layering and the plugin-extension guide.

License

MIT — see LICENSE.