Get started

ThunderPhone vs. building your own AI phone agent

Excellent open-source frameworks now exist for building voice agents, and some teams should absolutely use them. Building gives you maximum control: you choose every model, own every millisecond of the latency budget, and keep the whole stack inside your own infrastructure. Buying a finished platform is the opposite trade: a published per-minute price and someone else operating the pipeline while you focus on what the agent should say and do.

Every third-party fact below was checked against the cited framework's public documentation on August 19, 2026. The goal is not to talk anyone out of building — it is to make the full inventory of the build visible before the decision is made.

Decision summary: Build when the voice pipeline itself is your product or your differentiator — custom model routing, latency experiments, proprietary agent behavior — and you have engineers who will own it long-term. Buy when a phone agent is a means to an end: ThunderPhone's engines run at published all-in rates of 2¢, 5¢, or 9¢ per minute, with languages, telephony, transfers, knowledge, webhooks, and a testing suite already assembled.

What building your own actually requires

"The framework is open source" is true and also only the first line item. Here is the inventory a self-built phone agent needs, drawn from the frameworks' own documentation.

1. A framework and the agent code

The two leading open-source options are LiveKit Agents and Pipecat:

  • LiveKit Agents is a "Realtime framework for voice, video, and physical AI agents"; agents are Python or Node.js programs, the framework handles the STT–LLM–TTS pipeline, turn detection, and interruptions, and the ecosystem is Apache 2.0 licensed. Business logic is the developer's responsibility. LiveKit Agents docs, accessed 2026-08-19.
  • Pipecat is "an open source framework for voice and multimodal conversational AI," maintained by Daily and its community, BSD-2-Clause licensed, orchestrating 100+ AI services with client SDKs and a flow-management layer. Pipecat and GitHub repository, accessed 2026-08-19.

Both are genuinely capable. Either way, you write and maintain the agent application: prompts, tools, state, and error handling.

2. Model provider accounts and per-unit costs

Neither framework includes speech or language models in its license. Pipecat integrates 20+ STT, 30+ TTS, and 25+ LLM services, all bring-your-own-key — inference is "billed to provider." Pipecat GitHub and Pipecat Cloud pricing, accessed 2026-08-19. LiveKit's plugins likewise each require "your own account with the provider, as well as an API key or other credentials," with a managed LiveKit Inference path available on LiveKit Cloud. LiveKit models docs, accessed 2026-08-19.

Practically, that means three or four vendor accounts (STT, LLM, TTS, sometimes a realtime speech-to-speech API) — each with its own billing relationship, rate card, and status page.

3. Telephony, SIP trunking, and number management

A voice agent is not a phone agent until it can answer a phone. LiveKit supports bring-your-own third-party SIP trunks plus US numbers on LiveKit Cloud, configured through inbound/outbound trunks and dispatch rules. LiveKit SIP docs, accessed 2026-08-19. Pipecat's Twilio integration example requires a Twilio account, a voice-capable number, a TwiML Bin pointing at your bot's WebSocket endpoint, a publicly reachable server, and 8 kHz mono 16-bit PCM audio handling; its deployment docs flag production "SIP gotchas." Pipecat Twilio docs and deployment docs, accessed 2026-08-19.

You own number procurement, trunk configuration, webhook plumbing, and call-transfer behavior — including edge cases like voicemail detection and unanswered transfer targets.

4. Hosting and scaling

A Pipecat bot "is a Python process," and a production deployment needs the bot, an HTTP session-start service, and a media transport, with documented scaling patterns like VM-per-session or warm worker pools. Pipecat deployment docs, accessed 2026-08-19. Self-hosted LiveKit recommends Redis for production, 10Gbps-or-faster networking, compute-optimized instances, and leaves SSL certificates, UDP/firewall configuration, and public-IP handling to you. LiveKit self-hosting docs, accessed 2026-08-19.

Both projects offer managed clouds that remove much of this — LiveKit Cloud runs deployed agents with automatic scaling, and Pipecat Cloud scales bots from zero with CLI deploys. Even there, you still containerize, deploy, manage secrets, and monitor your own agent. LiveKit deployment docs and Pipecat Cloud docs, accessed 2026-08-19.

5. Conversation quality tuning

Turn-taking and interruptions are what make a phone agent feel usable, and the frameworks give you real machinery — LiveKit ships a custom turn-detection model; Pipecat documents a 500–800 ms conversational latency target. LiveKit Agents docs and Pipecat introduction, accessed 2026-08-19. Tuning it for your call patterns — barge-in sensitivity, silence thresholds, background noise — is iterative engineering work on real calls.

6. Monitoring, testing, and evals

LiveKit documents a real testing story: pytest/Vitest behavioral tests against an AgentSession in text mode, LLM-judged assertions, and cloud-run agent simulations — with the test cases and multi-turn flows authored by you. LiveKit testing docs, accessed 2026-08-19. For Pipecat, a built-in graded-scenario or regression-gate call-testing product was not established from the reviewed public sources; conversational testing is your own harness. Call logging, transcripts, alerting, and per-call debugging are also yours to wire up.

7. Ongoing maintenance

This is the line item that never appears on a pricing page. Providers ship new models and deprecate old ones; frameworks release breaking versions; telephony APIs change. A self-built agent is a living system with a permanent engineering owner, not a project with an end date.

What ThunderPhone provides as a finished product

ThunderPhone is the bought version of everything above, operated as one product:

  • Published all-in engine rates: pay as you go with no subscription or seat fee — Spark 2¢/minute, Bolt 5¢/minute, Storm 9¢/minute. Premium voice or language paths can add up to 3¢/minute; optional add-ons carry published surcharges (verbal acknowledgements +2¢ and call supervision +8¢ on Storm, long prompts +1–2¢ per additional 10K tokens, demo numbers +1¢/minute plus $1/month). Hold time bills a flat 2¢/minute; the builder displays the current all-in rate, and telephony and other optional configuration charges depend on setup.
  • Languages: 47 languages and 123 voices in the catalog, with automatic mid-call switching when a caller uses a configured additional language.
  • Telephony built in: inbound demo U.S. numbers for testing, plus production numbers via supported direct VoIP connections or manual SIP configuration for any SIP trunk; verified imported numbers support inbound and outbound calling. Cold and attended warm transfer are supported, including private briefing of the recipient and return-to-caller on decline.
  • Knowledge: a document library (text, Markdown, CSV, PDF, DOCX — 5 MB per text file, 50 MB per PDF/DOCX) plus URL import, with per-agent scoping and a mid-call knowledge search tool.
  • Integration: multiple webhook endpoints with HMAC-SHA256 signatures and exponential-backoff retries; remote MCP servers attachable by URL over streamable HTTP; an embeddable web widget with domain allowlists and origin-locked keys.
  • Testing as a product: browser mic tests, AI-caller simulations, bot-to-bot and SIP-loopback test calls, reusable graded scenarios, regression suites with minimum pass-rate gates, CI execution, and live-traffic A/B experiments. Simulations are billable real calls, with the charge shown before a run.
  • Compliance posture: GDPR and HIPAA compliant. Customers remain responsible for configuring their workflows, data handling, recordings, integrations, and notices for their own obligations.

The cost structure, honestly

The open-source software is free. The total is not: it is provider costs + infrastructure + engineering time, and no public page can compute it, because inference is bring-your-own-key on both frameworks.

What the public pages do quote are infrastructure floors. Using one illustrative workload — 2,000 calls × 5 minutes = 10,000 connected minutes per month — with every number below being arithmetic on published rates, not a quote:

  • Pipecat Cloud, smallest instance, PSTN telephony: hosting 10,000 × $0.01 = $100, PSTN dial-in/out 10,000 × $0.018 = $180 — a $280/month subtotal before any STT, TTS, or LLM usage (billed directly by your providers), recordings, noise suppression, or number rental. Pipecat Cloud pricing, accessed 2026-08-19.
  • LiveKit Cloud, Ship plan: $50/month including 5,000 agent session minutes, plus 5,000 overage minutes × $0.01 = $50, plus 5,000 third-party SIP minutes beyond the included 5,000 × $0.004 = $20 — a $120/month subtotal before model usage beyond the plan's $5 inference credit, phone-number rental where applicable, and your own SIP trunk provider's charges. LiveKit pricing, accessed 2026-08-19.

Those subtotals are floors, not totals. On top of them sit the model bills — not established from the reviewed public sources for any specific configuration — and the largest number of all: engineering time. As a labelled assumption rather than market data: if building and hardening a production phone agent absorbs even a fraction of one engineer's ongoing time, the loaded cost of that time will usually dwarf every infrastructure line item above. Run the calculation with your own salary numbers; the point is that it belongs in the model at all.

ThunderPhone, same workload: Spark 10,000 × $0.02 = $200/month, Bolt 10,000 × $0.05 = $500/month, Storm 10,000 × $0.09 = $900/month — illustrative base-engine arithmetic with no subscription fee. A premium voice or language path can add up to $300/month at this volume, and telephony and other optional configuration charges depend on setup. Only base engine usage is computable from minutes alone; the builder shows the configured all-in rate.

Build if / buy if

Build your own if:

  • The voice pipeline is itself your product, or agent behavior is your core IP — you need to own it outright.
  • You need custom pipeline stages, model routing, or providers a platform doesn't expose.
  • You are running latency experiments or research where controlling every pipeline stage matters more than time-to-launch.
  • You already have (or want to build) in-house voice-AI expertise, and a permanent owner for the stack is a feature, not a cost.
  • Your data or deployment constraints require fully self-hosted infrastructure end to end.

Buy ThunderPhone if:

  • Phone coverage is a means to an end — bookings, support, qualification, reception — and your engineers have a different job.
  • You want a bill you can forecast from minutes: 2¢/5¢/9¢ per minute, with surcharges published rather than assembled from three vendors' rate cards.
  • You need telephony, transfers, 47 languages, knowledge, webhooks, MCP, and a widget working this week, not this quarter.
  • You want testing — graded scenarios, regression gates, CI, A/B experiments — as a product feature instead of a harness you author and maintain.
  • You want GDPR and HIPAA compliance posture handled at the platform level, while you configure your own workflows, recordings, and notices for your obligations.

The middle ground — building on a framework but running it on the vendor's managed cloud — is covered in the dedicated pages: ThunderPhone vs. LiveKit and ThunderPhone vs. Pipecat.

FAQ

How much does it cost to build your own AI phone agent?

It cannot be computed from public pages, and that is the finding. Both LiveKit Agents and Pipecat leave STT, TTS, and LLM billing to your own provider accounts, so no published page states an all-in per-minute total. The quotable infrastructure floors for 10,000 monthly minutes are roughly $120/month (LiveKit Cloud Ship plan arithmetic) and $280/month (Pipecat Cloud smallest instance with PSTN) — before any model usage or engineering time. LiveKit pricing and Pipecat Cloud pricing, accessed 2026-08-19.

Is LiveKit Agents or Pipecat better for building a voice agent?

Both are credible, actively maintained open-source frameworks, and this page does not rank them. LiveKit Agents (Apache 2.0, Python/Node.js) ships turn detection, a plugin ecosystem, and a documented testing framework; Pipecat (BSD-2-Clause, Python) orchestrates 100+ services with client SDKs and a flows layer. LiveKit Agents docs and Pipecat docs, accessed 2026-08-19.

How long does it take to build a phone agent from scratch?

No credible fixed answer exists, so this page does not invent one. The frameworks' own docs enumerate the work: agent code, provider integration, telephony, deployment, turn-taking tuning, a testing harness, then ongoing maintenance as providers and models change. A demo can be quick; the gap between a demo and an agent that survives real callers is where the engineering time goes.

Can I start on a platform and move to self-built later, or vice versa?

Yes — the portable assets are your prompts, call flows, knowledge content, and test scenarios, not the pipeline. Neither direction is a lift-and-shift: telephony, webhooks, and testing are wired differently on each side. Well-documented prompts, scenarios, and grading criteria are what make either migration tractable.

Does buying a platform mean giving up SIP control or integrations?

Not on ThunderPhone. Production numbers can be brought through supported direct VoIP connections or manual SIP configuration for any SIP trunk, webhooks are signed with HMAC-SHA256 and delivered with retries, and remote MCP servers can be attached over streamable HTTP for tool integration. Demo numbers are inbound-only and not intended for production.

What is the real recurring cost difference?

Structure, more than any single number. Self-built: framework (free) + three or four provider bills + hosting + a permanent engineering owner, with only the infrastructure slice quotable in advance. ThunderPhone: published per-minute engine rates with published surcharges and no subscription or seat fee — $200–$900 of illustrative base-engine arithmetic at 10,000 monthly minutes, with telephony and optional configuration setup-dependent.

Sources and freshness

All LiveKit and Pipecat claims on this page were checked against the cited official pages on August 19, 2026; ThunderPhone pricing and product details reflect its published pricing and documentation as of the same day. This comparison is documentation-based — no framework deployment or live build was benchmarked — and all cost figures are illustrative arithmetic on published rates, not quotes; engineering-cost statements are labelled assumptions, not market data. Vendors change pricing and documentation often, so confirm against the live pages before relying on any figure here.