For agent builders

Give your agent a reason to pause.

Before an AI agent or a robot pays, opens or changes an account on a voice instruction, ORAVYS scores the recording for signs of synthetic speech. A risk signal before an irreversible action, never a lock.

Request API access → How it works

API v1 and MCP server
Consent required on every call
Paid API key, sent in a header
Recorded clips, not live streams

How it works

Record, score, then hold or step up.

The agent keeps doing the reversible work. Only the irreversible step waits for the result.

01 · Record with consent

The person is told, and agrees.

The agent announces the recording and gets consent from the person who speaks. Every call must attest that this consent was given. Without it the API refuses the call and counts nothing.

02 · Send the clip

One call, from your backend or through MCP.

One API call from your backend, or one tool call through our MCP server. For flows that can wait, the job runs in the background and your system checks back for the result.

03 · Hold or step up

The result decides the next check, not the action.

If the result is not clear, confirm through another channel before acting, for example a call back to a number you already know.

Timing: typically seconds when warm, longer on cold start; use the async job API for non-interactive flows.

Three possible answers

Each answer maps to a next step.

The result is advisory. It says whether synthetic speech signals were found in this recording, nothing more.

Answer

Strong synthetic signals

The recording shows strong signs of generated speech. Hold the irreversible action and confirm through another channel.

Answer

No strong synthetic signal (not a certification)

Nothing strong was found. This does not prove that the voice is human, nor who is speaking. Keep your usual controls.

Answer

Inconclusive

The clip could not be scored with confidence, or the service did not answer in time. Treat it as unknown, never as a pass.

What it is not

Not a lock.Never make it the only gate on an action. It tells your agent when to ask for more.
Not proof.A probabilistic output, not legal evidence and not an identity check. It does not tell who is speaking.
Not live call monitoring.It scores a recorded clip. There is no streaming endpoint.

Integration

The API you call, or the MCP server your agent uses.

One API call from your backend, or one tool call through our MCP server. Integration details are shared with API access. A key sent in the address or in a form field is refused. Keep the key on your server, never inside a device or a client app.

✓check_deepfake gives a short answer, no report. Limited to 10 calls per hour per key.
✓analyze_async submits a job and returns a job_id; get_job reads its status and result. Only the key that submitted a job can read it.
✓analyze_voice runs a full analysis in one call and can take minutes: not for an agent that is waiting.
✓The MCP server refuses locally, and sends nothing, unless consent is attested for both data processing and voice processing. It runs on your side over stdio; no remote MCP endpoint is offered.

Access

A paid key from the first call.

There is no free tier for automated agents: every analysis uses compute. Monthly quotas depend on the plan, and per-minute and per-day limits also apply. Access is opened on request.

Put a pause in front of the irreversible step.

Tell us what your agent does and which actions it takes on a voice instruction. We reply with access terms for the API and the MCP server.

Request API access →