resemble-detect

What it does

Analyzes audio, image, video, and text for synthetic manipulation, AI-generated content, watermarks, speaker identity, and media intelligence using the Resemble AI platform.

How it works

• Core rule: do not declare media real or fake until a completed detection result is returned, including label, score, and status: "completed". If processing, wait; if failed, report failure.
• Phase 1: Deepfake Detection – submit media to POST /detect with url and optional flags (visualize, intelligence, audio_source_tracing). Receive asynchronous results and poll GET /detect/{uuid} until status is "completed" or "failed". Interpret audio/video/image results from metrics or video_metrics with labels, scores, and visualization heatmaps.
• Phase 2: Intelligence – optionally run POST /intelligence to obtain structured results (speaker_info, language, emotion, transcription, etc.) for audio/video, or image-specific fields. Use POST /detects/{uuid}/intelligence to ask questions after a detection completes.
• Phase 3: Audio Source Tracing – if audio is labeled as fake, enable audio_source_tracing to identify the AI platform; results appear under audio_source_tracing.
• Phase 4: Watermarking – apply or detect watermarks via POST /watermark/apply or POST /watermark/detect. Apply supports image/video with optional strength and custom_message; use Prefer: wait for synchronous results or poll the result endpoint.
• Phase 5: Identity — Speaker Verification (Beta) – create identities with POST /identity and search with POST /identity/search to compare unknown audio against known profiles.
• Phase 6: Text Detection (Beta) – detect AI-generated vs human-written text via POST /text_detect, with blocking or asynchronous options using Prefer: wait.
• Media requirements: media must be publicly accessible via HTTPS URL; local files must be hosted publicly first.
• MCP tools exist for docs and API endpoints, invoked via resemble_docs_lookup, resemble_api_endpoint, etc., to fetch endpoint specifics.

When to use it

Use this skill when needing:

• AI-generated or manipulated media detection (audio, image, video, text)
• Media provenance and source tracing, watermarking, or watermark detection
• Speaker identity verification or intelligence extraction (emotion, transcription, misinfo)
• Questions about detection results after a completed detection
• Detecting AI-generated or machine-written text

What it can touch

• API endpoints: POST /detect, GET /detect/{uuid}, POST /intelligence, POST /detects/{detect_uuid}/intelligence, POST /watermark/apply, POST /watermark/detect, POST /identity, POST /identity/search, POST /text_detect
• Media must be provided as HTTPS URLs; local files must be hosted publicly first
• Tools via MCP: resemble_docs_lookup, resemble_search, resemble_api_endpoint, resemble_api_search, resemble_get_page, resemble_list_topics (used to fetch docs and exact schemas before calls)

Caveats

• Declares authenticity only after a completed detection result; do not rely on partial results
• Text detection and some features are Beta and may require specific roles or plans
• Beta features require appropriate access and may return limited availability or errors if access is not granted
• License notes in the skill metadata show Apache-2.0; the frontmatter in this instance indicates Apache-2.0