Agents, telephony, contact center, evals — one modern platform, end to end. You bring the models. We run everything else.
A scrap of sound, captured the instant it's spoken.
The line knows when the caller is done — and when they want back in.
Any ASR, any LLM, any TTS — wired as a DAG, configured per agent, swapped without a deploy.
OpenAI Realtime, Grok realtime, or Qwen Omni self-hosted on your GPUs.
First-class barge-in on every one, whichever model is behind it.
Explore in detail →Not just talk. Tools run alongside the audio, mid-call.
route_to_skill, transfer_call — blind, warm or consult — end_chat, present_form, plus any HTTP endpoint or MCP server you point them at.
The agent hands a live call to a human queue — hold music, claim, ring, bridge. Proven end to end on the public phone network.
Simulated callers, barge-in timing checks, LLM-judge grading, p50/p95 first-audio rollups — built into the platform.
Explore in detail →Carrier-grade SIP, engine-native — built from scratch, not a FreeSWITCH wrapper. Trunk in from Twilio, Vapi or straight from your carrier.
Announce, collect digits, check a skill, queue — the routing model contact centers already run, rebuilt AI-native.
Skill-based ACD with presence, five claim strategies, full ECMA CSTA CTI — ring in the browser, or on SIP desk phones if your floor runs them.
Live wallboards — service level, ASA, AHT, occupancy — and listen-in on any call, without touching it.
Explore in detail →One embeddable widget. The same agents, tools and routing that run your calls.
The agent pushes a form into the thread; the answers come back as a turn.
Capacity-aware routing offers chats on the same ring as calls — up to each agent’s limit.
Explore in detail →No LLM, no ASR, no TTS of ours. Your providers, your keys — sealed at rest, per tenant.
Every provider behind one alias table. Change a config, not your code — in production.
12 LLM providers, 9 streaming ASR, 5 streaming TTS, 3 realtime speech-to-speech — one integration.
One token off your own GPU, on your own terms.
Tokens leave the GPU and start arriving immediately.
Point your existing client at the edge. HTTP/3 halves WAN time-to-first-token versus HTTP/1.
Per-tenant model deployments on Kubernetes, metered and routed by the platform.
Explore in detail →Three commitments behind every layer.
No LLM, no ASR, no TTS of ours in the bill. Every provider runs on your keys, sealed at rest per tenant — swap any of them in production with a config change.
Engine-native SIP, browser-based agent audio, single-port sharding, CPU-only prompt rendering. No PBX licenses, no per-seat desk phones, no media servers to babysit.
The Avaya-shape routing your ops team knows — DIDs, vectors, skills, wallboards — rebuilt on QUIC and ClickHouse instead of twenty-year-old racks. And no FreeSWITCH or Asterisk under the hood: the media engine is ours.
Most voice AI platforms are a thin layer over FreeSWITCH, Asterisk or Pion. Our media engine is ours, top to bottom — same boxes, same workloads, measured:
Phone calls · G.711 answer + playback
Same load, half the CPU — head to head vs Pion
Same box, matched concurrent G.711 calls. ClutchCall scales roughly linearly; Pion climbs superlinearly — the gap widens as you load it.
Realtime media fan-out · CPU per viewer at 30 fps
About 750 calls per 1% of CPU — roughly 4× Pion, 7× Asterisk, 14× FreeSWITCH. And on realtime fan-out, 17× the viewers of Pion WebRTC.
Calls: concurrent G.711, identical hyperscaler VM; efficiency = sustained ÷ peak CPU. Pion ramp-ceilinged past 12,000 without saturating (62%); Asterisk setup-failed at 5,000; FreeSWITCH was CPU-bound at 4,500. Fan-out: 2-host GCE c4-highcpu-8, 1→N relay — 400 viewers / 2 Gbps on one core at 5.3% box CPU, 0% loss.
Engine-native failover, designed in from the start — not a load balancer bolted on. Active calls re-home to a surviving node with a sub-2-second media blip.
Every node owns one fixed RTP port per core, calls demuxed by 4-tuple. A call’s media home is just an address — re-homing it needs no per-call port allocation.
Heartbeats and a leader lease ride a dedicated Redpanda control plane, never the media path. Call ownership is registered the moment a call is established.
On answer, everything needed to reconstruct the call — dialog state, route, sequence, remote media, codec — is journaled to Redis, off-node.
The leader detects the dead node, hashes each call to a survivor, and the takeover node rebuilds the dialog from the journal and recovers it with an in-dialog re-INVITE.
Failover knowledge flows over Redpanda; media never waits on the control plane. Design target: <2 s media blip on node death, in-flight calls preserved.
Telequick is the platform underneath — not a walled garden on top. Run your agents on us end to end, or take transport only: your agents stay on LiveKit or Pipecat and ride our SIP core.
Bring your trunks: BYOC from Twilio or Vapi, or go straight to carrier SBCs. Your numbers land on engine-native SIP either way.
Point LiveKit agents at Telequick trunks and DIDs — transport-only is a first-class path. Your rooms get carrier-grade telephony, ACD escalation and recording underneath.
Run Pipecat pipelines against the same SIP core and provider catalog — agents stay in your DAG; you gain evals, routing and human handoff.
Browser audio first-class both ways: MoQT over WebTransport — uplink and downlink — where you have it, WebRTC where you don’t.
# a voice agent on the phone network, in 6 lines from telequick import Agent a = Agent(asr="deepgram", llm="groq/llama", tts="cartesia") a.tool("route_to_skill") # AI → human, native async for call in a.answer("+1..."): if call.bargein: a.cancel() # instant, every provider
Nine idiomatic SDKs, generated from one RPC schema — Python, TypeScript, Go, Rust and more. The voice events plane is included, so your app hears everything the platform hears.
Read the docs →You pay for lines, not minutes — because minutes should be free, and for us they basically are.
We’ll put it on the platform with you — your number, your providers, your keys — and you can judge the difference on a live call.
No. We sell no LLM, no ASR, no TTS. You bring provider keys — they’re sealed at rest, per tenant — or point us at models you host yourself. Our bill is for the platform, not for tokens.
Yes — two ways. Run Telequick as the full platform with your framework on top, or take transport only: your agents keep running on LiveKit or Pipecat while trunks, DIDs, routing, ACD, evals and recording ride our SIP core. Both paths are first-class.
Yes. Interruption detection lives in the platform — turn detection plus energy VAD — not in each provider’s SDK. Cancel fires the same way whether the agent is a cascaded pipeline or a realtime speech-to-speech model.
That’s the shape it’s built to. DID → VDN → vector routing, skill-based ACD with presence and claim strategies, CMS-style interval reporting, live wallboards and supervision — with AI agents and AI-to-human transfer native instead of bolted on.
Yes. The platform is multitenant SaaS by default, but the core — including self-hosted speech-to-speech inference on your GPUs — can run in your own racks.
No — voice minutes are unlimited. You pay for lines, not minutes, because minutes should be free, and for us they basically are: the media engine is cheap enough per call that metering minutes would cost more to run than the minutes are worth. You bring provider keys for models; the platform bill is capacity, not talk time.
We’re early, and we’d rather show you the platform than a rate card. Talk to us and we’ll scope it against what your current stack costs to operate.