Build a Bun + Hono + OpenAI Realtime Voice Agent on the Edge (2026)

TL;DR — Bun 1.3 starts in ~25ms cold, Hono is ~14kb, and OpenAI's gpt-realtime-2 (introduced 2026) gives you GPT-5-class reasoning over voice. Combined: a single bun run server.ts ships a 6-region edge voice agent.

What you'll build

A WebSocket relay between browser PCM and OpenAI Realtime, deployed to 6 Fly.io regions with anycast. Browsers get routed to the nearest edge, p95 voice-to-voice ~480ms.

Prerequisites

1. Bun 1.3+, hono@^4.6.
2. Fly.io CLI (brew install flyctl) and an OpenAI key.
3. Domain on Cloudflare for TLS pass-through.

Architecture

flowchart LR
  BR[Browser] --> CF[Cloudflare anycast]
  CF --> FY[Fly edge nearest of 6]
  FY -- WS --> H[Hono relay on Bun]
  H -- WS --> OA[OpenAI Realtime gpt-realtime-2]

Step 1 — server.ts

```ts import { Hono } from "hono"; import { upgradeWebSocket } from "hono/bun";

const app = new Hono(); const URL = "wss://api.openai.com/v1/realtime?model=gpt-realtime-2";

app.get("/ws", upgradeWebSocket(() => { let oa: WebSocket; return { onOpen: (_e, ws) => { oa = new WebSocket(URL, { headers: { Authorization: Bearer ${process.env.OPENAI_API_KEY}, "OpenAI-Beta": "realtime=v1" }, } as any); oa.onopen = () => oa.send(JSON.stringify({ type: "session.update", session: { voice: "verse", turn_detection: { type: "semantic_vad" } } })); oa.onmessage = (m) => ws.send(m.data); }, onMessage: (e) => oa?.readyState === 1 && oa.send(e.data), onClose: () => oa?.close(), }; }));

export default { port: 8787, fetch: app.fetch, websocket: { /* bun ws */ } }; ```

Step 2 — Dockerfile

```dockerfile FROM oven/bun:1.3-alpine WORKDIR /app COPY bun.lockb package.json ./ RUN bun install --frozen-lockfile COPY . . EXPOSE 8787 CMD ["bun", "run", "server.ts"] ```

Step 3 — fly.toml

```toml app = "voice-edge" primary_region = "iad" [build] [http_service] internal_port = 8787 force_https = true [[regions]] iad [deploy] strategy = "rolling" ```

fly deploy and fly regions add lhr nrt syd fra gru for 6 edges.

Step 4 — Browser PCM

Use a 24kHz AudioWorklet to capture PCM16 chunks every 20ms and forward as base64-wrapped input_audio_buffer.append events.

Step 5 — gpt-realtime-2 reasoning

The 2026 gpt-realtime-2 model handles complex multi-tool calls in one turn. Set session.instructions with up to 8K tokens of policy without hurting latency.

Step 6 — Anycast tip

Cloudflare's *.callsphere.ai → voice-edge.fly.dev with proxied = false (DNS-only) so WebSocket round-trips bypass the proxy.

Pitfalls

• Bun's WS spec drift: Some ws.send overloads differ from Node — test on Bun, not just locally.
• Fly cold start: Set min_machines_running = 1 per region to keep voice cold-starts <50ms.
• gpt-realtime-2 cost: Audio in/out roughly $0.07/$0.27 per minute (estimate; check current pricing).

How CallSphere does this in production

Explore a live demo and compare current plans to find the right fit for your business.

FAQ

Why not Node? Bun's WebSocket implementation is ~2x faster on raw throughput.

Cloudflare Workers? Workers cap WS connections at 6 hours and have no persistent state — Fly + Bun is simpler.

TURN servers? WebSocket relays don't need them; only WebRTC direct does.

Cost? Fly: ~$5/region/month for 256MB shared CPU. OpenAI: ~$0.20-$0.30/min.

Sources

• Hono on Bun WebSockets - https://hono.dev/helpers/websocket
• QuotyAI - Bun + Hono backend - https://quotyai.com/blog/why-i-picked-bun-and-hono/
• OpenAI gpt-realtime - https://openai.com/index/introducing-gpt-realtime/
• StartupHub - GPT-Realtime-2 2026 - https://www.startuphub.ai/ai-news/artificial-intelligence/2026/openai-s-new-voice-api-models

How this plays out in production

To make the framing in Build a Bun + Hono + OpenAI Realtime Voice Agent on the Edge (2026) operational, the trade-off you cannot defer is channel routing between voice and chat — a missed call should not die, it should warm up the SMS or web-chat lane within seconds. Treat this as a voice-first system from the first prompt: the agent's persona, its tool surface, and its escalation rules all flow from that single decision. Teams that ship fast tend to instrument the loop end-to-end before they tune any single component, because the bottleneck is rarely where intuition puts it.

Voice agent architecture, end to end

A production-grade voice stack at CallSphere stitches Twilio Programmable Voice (PSTN ingress, TwiML, bidirectional Media Streams) to a realtime reasoning layer — typically OpenAI Realtime or ElevenLabs Conversational AI — with sub-second response as a hard SLO. Anything north of one second of perceived silence and callers either repeat themselves or hang up; that single number drives the whole architecture. Server-side VAD with proper barge-in support is non-negotiable, otherwise the agent talks over the caller and the conversation collapses. Streaming TTS with phoneme-aligned interruption keeps the cadence natural even when the user changes their mind mid-sentence. Post-call, every transcript is run through a structured pipeline: sentiment, intent classification, lead score, escalation flag, and a normalized slot extraction (name, callback number, reason, urgency). For healthcare workloads, the BAA-covered storage path, audit logs, encryption-at-rest, and PHI-safe transcript redaction are wired in from day one, not bolted on at compliance review. The end state is a system where every call produces a row of structured data, not just a recording.

FAQ

What changes when you move a voice agent the way Build a Bun + Hono + OpenAI Realtime Voice Agent on the Edge (2026) describes?

Treat the architecture in this post as a starting point and instrument it before you tune it. The metrics that matter most early on are end-to-end latency (target < 1s for voice, < 3s for chat), barge-in correctness, tool-call success rate, and post-conversation lead score distribution. Optimize whatever the data flags as the bottleneck, not whatever feels slowest in your head.

Where does this break down for voice agent deployments at scale?

The two failure modes that bite hardest are silent context loss across multi-turn handoffs and tool calls that succeed in dev but get rate-limited in production. Both are solvable with a proper agent backplane that pins state to a session ID, retries with backoff, and writes every tool invocation to an audit log you can replay.

How does the After-Hours Escalation product make sure no urgent call is dropped?

It runs 7 agents on a Primary → Secondary → 6-fallback ladder with a 120-second ACK timeout per leg. If the primary on-call does not acknowledge inside the window, the next contact is paged automatically — voice, SMS, and push — until somebody owns the incident.

See it live

Book a 30-minute working session at calendly.com/sagar-callsphere/callsphere-llc-meeting and bring a real call flow — we will walk it through the live after-hours escalation product at escalation.callsphere.tech and show you exactly where the production wiring sits.