Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.masker.dev/llms.txt

Use this file to discover all available pages before exploring further.

This guide walks you through connecting a Vapi assistant to Masker and placing a real call. By the end, you’ll have a working masked agent and a session report you can download — and you’ll be able to confirm in OpenAI’s own completion logs that no raw PHI reached the model.
You need a Vapi account with a phone number or web call testing enabled, an OpenAI API key, and a GitHub account for Masker portal sign-in.

​
Steps

1

Sign in to the Masker portal

Open masker-voice.fly.dev/portal/login and click Sign in with GitHub. Masker uses GitHub OAuth and stores nothing from your GitHub profile beyond your email address.After sign-in you’ll land on the dashboard with a prompt to connect your first agent.
2

Create an agent

Click New Agent on the dashboard. The setup wizard has three fields:
  • Agent name — use something descriptive, like Patient intake (test) or Prior auth — primary care. This name shows up in your session list.
  • Vapi API key — copy your key from dashboard.vapi.ai/account and paste it here. Masker uses the key once to configure the assistant on Vapi’s side, then stores it encrypted using AES-256-GCM-SIV.
  • Upstream model — defaults to gpt-4o-mini. Change this if you want a different OpenAI model.
Use your own dev Vapi account for initial testing, not a customer’s key. The API key has full permissions on your Vapi org.
3

Copy your proxy URLs

After creating the agent, Masker generates two URLs unique to this agent:
  • Proxy URL — https://masker-voice.fly.dev/proxy/{agent_id}/v1/chat/completions
  • Webhook URL — https://masker-voice.fly.dev/vapi/webhook/{agent_id}
Masker configures these on your Vapi assistant automatically. You don’t need to paste them anywhere for Vapi — they’re shown for reference and for non-Vapi platforms.The agent appears on the dashboard with a green Live badge once setup completes.
4

Verify the Vapi assistant config

Open your Vapi dashboard → Assistants and find the assistant Masker just created (it uses the same name as your agent). Confirm:
  • Model provider: Custom LLM
  • Custom LLM URL: your Masker proxy URL
If you want to route an existing Vapi assistant through Masker instead, go to that assistant’s settings, switch the model provider to Custom LLM, and paste the proxy URL.
5

Place a test call

Use the Talk to Assistant button on the Vapi assistant detail page, or call the phone number assigned to the assistant.Try saying something with clear PHI:
“Hi, my name is Sarah Johnson. My date of birth is March 14, 1982, and my member ID is BCBS-447299.”
Your assistant should respond naturally. Behind the scenes, Masker:
  • Tokenized Sarah Johnson → MSKV1.person_name.K_HEALTHCARE.a3f9...
  • Tokenized March 14, 1982 → MSKV1.dob.K_HEALTHCARE.b7c2...
  • Tokenized BCBS-447299 → MSKV1.account_number.K_HEALTHCARE.c1d8...
  • Forwarded the masked prompt to OpenAI
  • Rehydrated the tokens in OpenAI’s response before returning it to Vapi
6

Verify in OpenAI's logs

Open platform.openai.com/logs and find the chat completion that just ran. The Input section should look like this — no raw names, dates, or member IDs anywhere:
User: Hi, my name is MSKV1.person_name.K_HEALTHCARE.a3f9...
      My date of birth is MSKV1.dob.K_HEALTHCARE.b7c2...
      and my member ID is MSKV1.account_number.K_HEALTHCARE.c1d8...
This is the proof your compliance story rests on. OpenAI keeps completion logs for 30 days by default — even on enterprise plans. Masker ensures those logs never contain real PHI.
7

Download the session report

Back in the Masker portal:
  1. Open Sessions and click the session that just completed.
  2. Toggle between Raw and Masked views to see the transcript split across the compliance firewall.
  3. Click Download report to get the signed compliance report.
The report includes the BAA chain (Vapi → OpenAI → Deepgram → ElevenLabs), the HIPAA Safe Harbor coverage breakdown for this session, and a count of each PHI category detected.

​
What’s next

How it works

The full request pipeline — detection, tokenization, proxy forward, rehydration, and audit chain.

Request beta access

For production use with a signed BAA and SLAs — May 30, 2026 launch.