Skip to main content
Tuteliq is more than a detection API. Every detection can feed an optional moderation layer: a queue of incidents, triage tools, human and AI agent review, and a signed record of every decision. It runs on the same models as the API, and it never stores the content it judged. You choose how much of it to use:

Detection only

Call the API and act on the response in your own systems. Turn incident logging off and Tuteliq keeps no record at all. See Incident Logging.

Detection + moderation

Leave incident logging on and work the queue in the Tuteliq dashboard, over the incidents API, or with an AI agent over MCP.

The moderation loop

1

Detect

A detection call returns the verdict, severity, rationale and recommended action in the response, as always.
2

Log an incident (optional)

When a detection warrants attention, Tuteliq records an incident: the model’s assessment, never the content. Summaries are encrypted at rest, or end-to-end encrypted if you register your own key. See what an incident holds.
3

Triage the queue

Size the backlog, filter by status, severity, category, source or your own identifiers, and work the most urgent first. Every incident carries its recommended actions, so the queue can be ranked without opening each one.
4

Review and decide

A moderator confirms, downgrades, reclassifies, dismisses or escalates. An AI agent may only confirm a detection; anything that changes or overrides it is reserved for a person.
5

Keep the evidence

Every decision emits its own signed audit receipt, including each decision in a batch. Trends, overviews, action plans and incident reports turn the record into evidence for a regulator, a school counsellor or an external hand-off.

Three ways to work the queue

Incidents API

MCP tools for moderation

moderation_queue, list_incidents, get_incident, review_incident, batch_review_incidents, get_incidents_overview and get_incident_trends. See Moderation Workflows with an AI Agent for how an agent works a queue and what it is allowed to decide.

Human oversight by design

An MCP token identifies an account, never an individual, so an agent cannot act with a named moderator’s authority. Agents can confirm a detection. Downgrading, reclassifying, dismissing and escalating are for people, which keeps a human in the loop for every decision that overrides the model, in line with the human-oversight requirement of EU AI Act Article 14.

Privacy, with the moderation layer on

  • No content, ever. An incident holds the model’s assessment: category, severity, confidence, rationale and recommended actions. Never the message, image, audio or video.
  • Encryption. Summaries are encrypted at rest by default, and end-to-end encrypted if you register your own public key, so Tuteliq cannot read them.
  • Retention you set. 90 days by default, anywhere from 7 days to 2 years with incident_retention_days.
  • Off at any time. incident_moderation_enabled: false stops incident records entirely; detection keeps working in full.

For hotlines and high-sensitivity teams

Teams that must keep material off any vendor’s servers can still use the moderation tools:
  1. Send only what detection needs, and keep the original material in your own systems.
  2. Register your own key, so incident summaries are end-to-end encrypted and readable only by you.
  3. Set a short retention period, such as 7 or 30 days, long enough to work the queue.
  4. Record decisions through the dashboard, the incidents API or MCP, and keep the signed receipts as your evidence.
Or turn incident logging off and use detection alone, keeping every record on your side.