QuietBridge

QuietBridge is a context-aware privacy control plane for an Alexa+ shared-space experience. It classifies returned fields, evaluates the current audience, routes only safe content to public speech, creates a short-lived private handoff, explains the privacy decision, records a redacted receipt, and reports deletion outcomes honestly.

The product is connector- and channel-independent: an existing calendar, health, commerce, or account service plugs into the server-side connector contract while Alexa+, another assistant, or an application consumes the real MCP interface. The included integration workspace uses invented Alexa, identity, phone, and calendar adapters to demonstrate the complete product without customer data.

Submission evidence

• Devpost submission copy

• Three-minute demo script

• Submission checklist

• Verification matrix

The second captured flow shows the product choosing a private handoff instead of speaking protected values in a shared room:

Related MCP server: privacy-gateway-mcp

Requirements

• Node.js 22 or later

• npm 11 or later

Run

npm install
npm run quality
npm run test:e2e
npm run dev

Open http://127.0.0.1:5173. The development command runs both the MCP server and integration workspace. Use npm start when only the development MCP server is needed. npm run build compiles both the deployable server and the lazy-loaded production UI; npm start:prod runs the compiled server.

The MCP endpoint is http://localhost:3100/mcp. The server supports the modern 2026-07-28 MCP protocol and 2025-era Streamable HTTP compatibility through the official SDK handler.

Optional AWS Bedrock classifier

QuietBridge can use Amazon Bedrock to classify the metadata of fields that are not in its deterministic sensitivity map. The model receives the field path, source name, and record type only—never the field value. The policy engine remains authoritative: Bedrock cannot downgrade an unknown field to public, and any error or low-confidence result remains fail-closed as personal data.

See docs/aws-bedrock-setup.md before enabling it. Bedrock inference is usage-priced and is disabled in the free-only configuration.

An optional Amazon Verified Permissions adapter can provide a second, centralized authorization check for protected results. It receives an opaque SHA-256 profile identifier and audience metadata, not protected values. A remote allow never relaxes the deterministic local decision; deny, error, or invalid response blocks the request. See docs/aws-verified-permissions.md. Verified Permissions is also disabled in the free-only configuration.

Free encrypted local handoff storage

By default, private handoffs remain in memory. For restart-safe local demos without paid services, QuietBridge also includes an AES-256-GCM encrypted file store. It persists token hashes rather than raw tokens, filters authentication secrets before storage, detects tampering, and preserves one-time consumption and revocation state.

See docs/local-encrypted-storage.md. Keep the key outside source control; .env, .quietbridge/, and *.enc.json are ignored.

Privacy guarantees under test

1. Sensitive values never enter public output.

2. Unknown classification fails closed.

3. A private handoff is profile-bound, short-lived, and one-time.

4. Explanations and receipts contain categories and reasons, not hidden values.

5. Deletion status distinguishes local success from connector success.

6. Optional model assistance never receives protected field values and can only strengthen, not weaken, an unknown field's category.

7. Browser flows, deletion, MCP evidence, and serious/critical accessibility violations are covered by Playwright and axe.

Productization and integration

• Product architecture and technical differentiation

• Connector, identity, channel, and container integration

• Security controls and measurable quality gates

• Three-minute working-product demonstration

Production mode is deliberately fail-closed. It requires authenticated MCP/private access, encrypted handoff persistence, trusted hosts, and allowed origins at startup. The repository is publishable as an open-source product and integration reference; the remaining real-customer deployment responsibilities are stated explicitly rather than hidden behind demo claims.

See quietbridge-build-and-test-plan.md for the complete architecture and submission plan.

Current implementation truth, remaining limitations, and the verification evidence are in:

• docs/implementation-status.md

• docs/verification-matrix.md

• docs/friction-log.md