Tuteliq MCP

Tuteliq MCP Server

MCP server for Tuteliq - AI-powered child safety tools for Claude

API Docs • Dashboard • Trust • Changelog • Discord

---

What is this?

Tuteliq MCP Server brings AI-powered child safety tools directly into Claude, Cursor, and other MCP-compatible AI assistants. Ask Claude to check messages for bullying, detect grooming patterns, or generate safety action plans.

Reads context, not just keywords. Every detector understands coded slang, emoji, leetspeak, algospeak, and deliberate filter evasion, and weighs the conversation around a message — so it tells gaming trash-talk apart from targeted harassment instead of drowning your team in false positives. This coded-language resilience is platform-wide (it applies to grooming, fraud, radicalisation, and the rest, not just bullying) and is built from our own research into how bad actors evade moderation. In an internal benchmark of coded-language and filter-evasion cases, Tuteliq detected roughly 1.7x more of them than leading general-purpose moderation APIs (319-case evasion set; vendors unnamed).

Fast mode. Pass verdict_only: true on detect_grooming or detect_bullying to get just the verdict (risk level, flags, recommended action) without the per-message breakdown — lower latency for real-time screening. The verdict itself is unchanged.

Interactive results. In hosts that support MCP Apps, results render as interactive cards rather than walls of JSON — see Interactive widgets below.

Interactive widgets

Twelve widgets return a rendered card instead of raw text in hosts that support MCP Apps (Claude desktop and web, and other MCP-compatible clients). Everywhere else the same data arrives as structuredContent, so nothing depends on the UI.

Every card carries the same frame: a chrome bar naming the tool that produced the result, the result itself, and a footer with the data-handling note and a Trust Center link. In a transcript holding a dozen results, the chrome bar is what tells you which is which.

Severity is rankable by colour. The ramp runs monotonically from safe to critical, so two chips can be compared without reading their labels:

Design notes. The widgets are deliberately calm. They report on grooming, self-harm, and abuse, and a card that animates or pulses at the reader turns material that is already distressing into an alarm they cannot dismiss. Severity is carried by a rule and a glyph, not by motion. The crisis-support card leads with reassurance rather than the severity colour, and its helpline numbers are the largest targets on the card because transcribing digits under stress is where people fail.

Widgets are read-only renderers by design. Selecting incidents in the list widget assembles the ID list for a batch_review_incidents call you fire yourself — the mutating call still goes through your host's approval step, so the human-in-the-loop stays in the loop.

Working on the widgets

npm run preview:ui   # builds every widget against fixture data

open dist-preview/__preview.html

Widget source lives in ui/src. Design tokens are centralised in ui/src/theme.ts; prefer them over colour literals so the palette stays in one place.

Available Tools (81 MCP)

Safety Detection

Fraud & Harm Detection

Voice, Image, Video & Document Analysis

Synthetic Content Detection

Identity & Age Verification

These three tools are registered when you run this package yourself. The hosted connector does not serve them.

Incidents & Moderation

Read the incident store, triage a queue, and record moderator decisions. The review tools emit signed receipts for EU AI Act Art 14 human-oversight evidence.

The decision is the moderator's, and the card makes them take it. The action buttons call review_incident through the host, because a moderator clicking "Escalate" *is* the human decision. They do not fire on one click: review_incident persists an override and emits a signed Art 12 audit receipt and requires a reason_code, so the button opens a reason picker and a second click commits. Nothing is ever defaulted into that receipt on the moderator's behalf.

The reasoning, confidence and analysis trace on the card are supplied by the calling assistant and are labelled as such — an argument for a human to weigh, not a Tuteliq measurement.

Pass operator_name to brand the header with the customer or team name. Omit it and the card is unbranded — it is never defaulted to a placeholder.

Webhook Management

Pricing

Usage & Billing

Policy Configuratio