@memberjunction/ai-bridge-twilio

The Twilio Realtime Bridge driver — the first telephony bridge in MemberJunction's Realtime Bridges program, and the reference driver the RingCentral and Vonage telephony drivers copy. It connects the one realtime agent engine to a phone call over Twilio Programmable Voice + Media Streams: outbound dial, inbound DID routing, single-party audio in/out, DTMF send/receive, and call transfer — all behind an injectable telephony call SDK seam so the driver builds and unit-tests with no network and no real Twilio client.

A phone call goes through the same transport seam as every meeting bridge — it is just a single-leg audio media source. Telephony is audio only: no video, no screen, and no Meeting Controls / facilitator surface (a 1:1 call has no roster to facilitate).

See the Realtime Bridges Guide (§ Telephony bridges) and /plans/realtime/realtime-bridges-architecture.md (§8 Twilio capability row, §9 Phase 6) for the full architecture.

Install

npm install @memberjunction/ai-bridge-twilio
Copy

What it provides

• TwilioBridge — @RegisterClass(BaseRealtimeBridge, 'TwilioBridge'). The MJ: AI Bridge Providers row with DriverClass = 'TwilioBridge' resolves to this driver via the ClassFactory. It is a thin subclass of BaseTelephonyBridge — all call lifecycle, the audio media seam, DTMF, transfer, the caller+agent roster, and capability gating are inherited. The driver only binds the Twilio SDK factory.
• TwilioCallSdk — the Twilio binding of the platform-agnostic ITelephonyCallSdk seam over Twilio Programmable Voice + Media Streams. Ships unbound (every op throws "bind the real Twilio client") until a deployment supplies ITwilioClientBindings over the real twilio SDK.
• TwilioCallSdkFactory — the creation seam that builds a TwilioCallSdk from resolved config.

Capability coverage (the Twilio seed row)

Capability gating is two-layer (defense-in-depth): the engine checks the provider's SupportedFeatures first, and the base re-asserts each flag with RequireFeature at the top of its overrides.

Outbound vs inbound

The shipped RealtimeBridgeContext is a meeting-shaped contract with no Direction field, so the engine forwards the session's direction (and the agent's caller-id / the inbound call id) into the driver's Configuration:

Connect requires AudioIn + AudioOut; outbound additionally requires OutboundDial, inbound requires InboundRouting.

The Twilio binding (deployment TODO)

TwilioCallSdk maps the telephony seam onto Twilio's two halves — the REST API (place/modify/hang-up) and the Media Streams websocket (bidirectional realtime audio + DTMF):

Out of the box TwilioCallSdk is unbound — every operation throws an explicit "bind the real Twilio client" error. Bind the real client by supplying an ITwilioClientBindings (REST + Media Streams) when constructing the SDK; the driver and its tests do not change, and none of the twilio SDK's types leak into this package. Credentials (Account SID, Auth Token, Media-Streams URL) resolve through MJ's credential system referenced by the provider Configuration — never inline secrets.

Usage (engine-driven)

The bridge is not used directly — AIBridgeEngine.StartBridgeSession (@memberjunction/ai-bridge-server) resolves it from the provider's DriverClass, forwards the session Direction / caller-id into the driver config, and wires the transport seam to the injected IRealtimeSession. There is no channel host to wire for telephony (no Meeting Controls).

Testing

FakeTwilioCallSdk (in src/__tests__/) is an in-memory ITelephonyCallSdk with drive helpers and capture sinks. The suite covers outbound dial → connect → audio round-trip (in/out), inbound answer, DTMF send + receive, transfer (gated), hangup + onCallEnded, the single caller+agent roster, capability gating (video/screen/Meeting Controls correctly absent), and the unbound-TwilioCallSdk bind-me throw — all with no network.

cd packages/AI/Providers/BridgeTwilio && npm run test
Copy

Production REST wiring (RealTwilioRestClient)

RealTwilioRestClient is the production ITwilioRestLike over the real twilio SDK (an optionalDependency, lazily loaded — the package still builds/tests with no twilio install). It maps CreateCall → client.calls.create({ to, from, twiml, statusCallback }) and UpdateCall → client.calls(sid).update({ status, twiml }). Construct it with credentials resolved from MJ config (API-key pair preferred over the account auth token); pass it as the Rest half of RealTwilioBindings (the MediaPump half is the Media-Streams websocket, owned by the MJAPI telephony ingress). Unit-tested via an injected fake twilio factory — no network.

Credential-gated integration test (NOT in CI)

real-twilio-bindings.integration.test.ts places a real outbound call and ends it. It is skipIf-gated on these env vars and never runs in CI (it spends real money + dials real PSTN):

Run locally: npm run test -- real-twilio-bindings.integration. The full media round-trip (audio through the agent) needs the MJAPI ingress + a publicly reachable stream URL — see the T1 ingress notes + manual runbook.