InNetwork: A Doctor for Each Family Member, in Their Own Plan's Network, by Voice
Built in-house
Retailbox R&D: built in-house for the Amazon Build, Ship, Shape hackathon (Alexa+ track), 2026
A voice add-on for Alexa+: say one sentence, and it finds a doctor for each person in the family, in their own plan's network, who sees a patient that age, and books around the family's day.
What was the problem?
A family on more than one health plan finds out by phone whether a doctor takes each person's plan and sees a patient that age. The public data doesn't help much. An audit of six public insurer directories done for this project found that 0 of 10,075 records said what age a doctor sees.
How did Retailbox solve it?
InNetwork is an Alexa+ add-on built as a self-hosted MCP server, with MCP Apps screens for Echo Show. One sentence, such as 'a gastroenterologist, an allergist for my five-year-old, and an orthopedist for my little one, he's under two', is split into three people, three plans and three searches. Each person is checked against their own plan's network, using an insurer's FHIR Plan-Net directory, the CMS marketplace provider file and a Medicaid managed-care directory. A deterministic gate sets aside doctors who can't see that person and says why, and the source of each reason is kept with it. Then the visits are placed around the family's routine. The checkup year is remembered across sessions, so 'what's left?' still works the next day.
No AI model runs inside the product. It has 13 refusals written as code, each with a test: it doesn't pick a specialty or read symptoms, won't book out of network without a second yes, doesn't store a child's age and doesn't read member IDs aloud.
The Alexa+ add-on toolkit is in preview with select partners, so we simulate Alexa+ in a web host that ships with the project. The doctors on screen are fictional, every name checked against the public NPPES registry so it matches nobody. Booking and office age policies in the demo are sandbox data and labelled that way. A classic Alexa skill that calls the same server ran on a real Echo Show 8 on 20 October 2026. The audit code is also released as a separate open-source package, plannet-directory-check (MIT).
Facing the same problem? These are the services that cover this kind of work: AI agent development services.
What does it look like?
(opens the full-size image in a new tab)
(opens the full-size image in a new tab)What were the results?
- ✓0 of 10,075 public directory records (5,649 practitioner roles and 4,426 locations, six insurers) state a patient age; one command re-runs the audit, with a control record that proves the counter works
- ✓13 refusals in code; a mutation check removes each rule in a throwaway copy and proves its test fails: 16 of 16 caught
- ✓68 of 68 verify checks pass on a fresh clone in about 45 seconds, offline except one check against the public NPPES registry
- ✓9 typed MCP tools with strict input schemas, checked independently with MCP Inspector in strict mode