The detectors weigh the surrounding conversation and are built for coded slang, emoji and deliberate filter evasion, which is what separates gaming trash-talk from targeted harassment. The half people underestimate is the incident side: review_incident requires a reason code and emits a signed receipt, so the human decision is recorded as one rather than inferred.
A hosted MCP server for Tuteliq's child-safety and content-risk detection API. It scores text, voice, images, video and PDFs for harm across twelve detection types, produces age-appropriate action plans and incident reports, and carries the whole moderation loop — queue, decision, signed receipt — behind the same connection.
- detect_bullying, detect_grooming, detect_unsafe, analyze and analyse_multi — harassment and gaming toxicity, predatory patterns, self-harm and violence; analyse_multi runs up to ten detection endpoints over one piece of text in a single call
- batch_analyze and analyze_emotions — up to 50 items across all twelve detection types in one request, and emotional and mental-state indicators over a message or a full conversation
- detect_social_engineering, detect_app_fraud, detect_romance_scam, detect_mule_recruitment, detect_gambling_harm, detect_coercive_control, detect_vulnerability_exploitation and detect_radicalisation — the fraud and harm classifiers
- analyze_voice, analyze_image, analyze_video and analyze_document — transcribe and analyse audio, OCR and inspect images, sample video key frames, and run per-page detection over a PDF with chain-of-custody hashing
- detect_synthetic_text, detect_synthetic_image, detect_synthetic_audio and detect_synthetic_video — AI-authorship forensics, including EXIF, C2PA Content Credentials, watermarks and lip-sync correlation
- get_action_plan and generate_report — age-appropriate guidance for a situation, and an incident report built from the conversation
- get_incidents_overview, list_incidents, get_incident, get_incident_trends and moderation_queue — counts by category and severity, a filterable list, one incident's risk trajectory, and the unreviewed triage queue
- review_incident, batch_review_incidents, get_audit_receipt and get_audit_logs — record a moderator decision with a reason code and a signed receipt, apply one decision across many incidents, and query the trail
- create_verification_session, get_verification_session and cancel_verification_session — age or identity verification with document intelligence, face match and liveness
- get_policy, set_policy and the policy-rule tools — per-category flag and block thresholds, and rules that act automatically when a detection matches, dry-runnable before you enable them
- The webhook, usage and threat-intelligence tools — endpoints and their signing secrets, per-tool and per-period usage, real-time quota, and anonymised network-wide threat trends
A Tuteliq account. The recommended path is no install at all: point the client at https://api.tuteliq.ai/mcp and sign in through the browser — it implements OAuth 2.1 with dynamic client registration and PKCE, so nothing is pasted into a config file and access is revoked from the dashboard. Headless callers send a static token in the Authorization header instead. Clients that only speak stdio can run npx -y @tuteliq/mcp with TUTELIQ_API_KEY; Node.js 18 or newer.
One command plus a key — npx -y mcp-remote https://api.tuteliq.ai/mcp, then supply credentials