# Product recipe: Voice agent (phone or web) (AgentGid)

For an assistant people talk to: speech in, a model in the middle, speech out, on the phone or in the browser.

Parts were picked from AgentGid's catalog by rules: free tier, official MCP server, use in public projects built with AI agents, downloads, price. Prices are list prices from the official pages (see each card for the check date); confirm before you commit.

## What it costs
- Start: these first picks have no free tier listed: Twilio (voice and SMS).
- If you outgrow the free tiers, the lowest paid plans of the first picks are Cartesia $5, Cal.com $12, LangSmith $39 (together about $56/month); usage beyond a plan's credits is billed on top.
- Billed by usage (grows with traffic): Deepgram, Twilio (voice and SMS).
- One minute of a call: about $0.036 (≈ $36 per 1,000 minutes) = speech-to-text Deepgram $0.0077 + voice (text-to-speech) Cartesia $0.0188 ($0.0376/min × 50%) + phone line (receiving) Twilio (voice and SMS) $0.0085 + model Claude Haiku 5.5 $0.0013. AgentGid estimate; assumes the agent talks 50% of each minute, about 1,000 characters of speech a minute, 4 model turns a minute at ~2,500 input + 150 output tokens.
- All-in-one platforms, per minute as listed: Retell AI from $0.07; Vapi (voice agent API) $0.05 platform fee + providers; LiveKit $0.01 platform fee + providers; Ultravox from $0.05.
- Model, default: fast and cheap: about $0.33 per 1,000 conversation turns with Claude Haiku 5.5 (AgentGid estimate at ~2,500 input + 150 output tokens each).
- Model, harder questions: about $2.44 per 1,000 conversation turns with Gemini 3.8 Flash (AgentGid estimate at ~2,500 input + 150 output tokens each).
- Prices of the picks below were checked on the official pages between 2026-10-09 and 2026-10-22.

## How it fits together
1. The phone provider (or the platform) receives the call and streams the audio to your server, or you use one all-in-one platform for the whole call.
2. Speech-to-text turns the caller's speech into text as they talk.
3. The model answers and calls your tools: check free slots and book through the calendar API.
4. Text-to-speech speaks the answer back over the same call.
5. Log each call (transcript, tool calls) and hand over to a person when the agent is unsure.

## Speech-to-text
- Deepgram: free: $200 credit, no credit card · usage: Nova-3 streaming $0.0048/min (promo; regular $0.0077); pre-recorded $0.0043/min; Flux $0.0065/min (promo); Aura-2 TTS $0.030 per 1k characters · official MCP · npm @deepgram/sdk · pip deepgram-sdk · card https://agentgid.com/apis/deepgram/agent.md
- AssemblyAI: free: Up to 185 hours pre-recorded and 333 hours streaming transcription, no credit card · usage: Universal-2 $0.15/hour; Universal-3.5 Pro $0.21/hour; Universal-Streaming $0.15/hour; diarization +$0.02/hour · official MCP https://assemblyai.com/docs/mcp · npm assemblyai · pip assemblyai · card https://agentgid.com/apis/assemblyai/agent.md
- Azure AI Speech (speech-to-text): free: Real-time Transcription: 5 audio hours free per month · usage: Standard: real-time $1.00 per hour, batch $0.18 per hour, fast transcription $0.36 per hour; custom real-time $1.20 per hour (East US) · official MCP @azure/mcp · npm microsoft-cognitiveservices-speech-sdk · pip azure-cognitiveservices-speech · card https://agentgid.com/apis/azure-speech-to-text/agent.md · MCP server card https://agentgid.com/mcp/servers/com-microsoft-azure/agent.md
- Also fits: Cartesia, Whisper (self-hosted), faster-whisper, ElevenLabs (https://agentgid.com/apis/speech-to-text/)

## Voice and real-time audio
- Cartesia: free: 20K credits/month + $1 prepaid for agents · paid from $5/mo · usage: Pro $5/month for 100K credits (~133 TTS minutes); Managed Agents $0.06/min; telephony $0.014/min · official MCP cartesia-mcp · npm @cartesia/cartesia-js · pip cartesia · card https://agentgid.com/apis/cartesia/agent.md
- Fish Audio API: free: s2.1-pro-free model at $0.00 / M UTF-8 bytes · usage: TTS $15.00 / M UTF-8 bytes (s2.1-pro, s2-pro); STT $0.36 / audio hour (transcribe-1-pro); voice design $0.01 / successful request · official MCP https://docs.fish.audio/overview/mcp · npm fish-audio · pip fish-audio-sdk · card https://agentgid.com/apis/fish-audio/agent.md
- ElevenLabs: free: Free plan: 10,000 TTS characters/month (v4) · paid from $6/mo · usage: TTS v4 $0.08 per 1K characters; Flash/Turbo $0.04 per 1K; Scribe v2 STT $0.22/hour; realtime $0.39/hour · official MCP elevenlabs-mcp · npm @elevenlabs/elevenlabs-js · pip elevenlabs · card https://agentgid.com/apis/elevenlabs/agent.md
- Also fits: Daily, OpenAI Realtime API, Rime, Typecast TTS API (https://agentgid.com/apis/voice/)

## Phone numbers and calls
- Twilio (voice and SMS): no free tier listed · usage: US: calls $0.0140/min out, $0.0085/min in (local number $1.15/mo); SMS $0.0083 per segment · npm twilio · pip twilio · card https://agentgid.com/apis/twilio/agent.md
- Bandwidth: no free tier listed · usage: US 10DLC SMS $0.004 per message, MMS $0.015; toll-free SMS $0.007; short code SMS $0.008; US local outbound voice $0.0100 per minute, inbound $0.0055 per minute · official MCP https://dev.bandwidth.com/docs/mcp · npm bandwidth-sdk · pip bandwidth-sdk · card https://agentgid.com/apis/bandwidth/agent.md
- Also fits: Infobip, MessageBird, Plivo (https://agentgid.com/apis/messaging/)

## Or one voice-agent platform (all in one, per minute)
- Retell AI: free: $10 in free credits · usage: $0.07-$0.31/min for AI voice agents; $0.002+/msg for chat agents · official MCP @retell-ai/mcp-server · npm retell-sdk · pip retell-sdk · card https://agentgid.com/apis/retell-ai/agent.md
- Vapi (voice agent API): free: $5 free credits to start · paid from $29/mo · usage: $0.05/min Vapi hosting; models, speech and telephony billed at cost on top · official MCP @vapi-ai/mcp-server · npm @vapi-ai/server-sdk · pip vapi_server_sdk · card https://agentgid.com/apis/vapi-api/agent.md
- LiveKit: free: Build plan: 1,000 agent session minutes, 1 free US phone number, inference credits · paid from $50/mo · usage: $0.01 per agent session minute, $0.01 per inbound calling minute, model-based inference pricing · npm livekit-client · card https://agentgid.com/apis/livekit/agent.md
- Also fits: Ultravox (https://agentgid.com/apis/voice/)

## Calendar and booking
- Cal.com: free: Free forever, 1 user, unlimited event types & calendars · paid from $12/mo · npm @calcom/embed-react · card https://agentgid.com/apis/cal-com/agent.md
- Also fits: Calendly, Cronofy (https://agentgid.com/apis/business-apps/)

## Traces and evals
- LangSmith: free: Developer plan: 1 seat, up to 5k base traces / mo, then pay-as-you-go · paid from $39/mo · usage: Plus: $39 / seat per month incl. up to 10k base traces / mo; trace overage is pay-as-you-go (rate not listed on pricing page); 1 LSU = $1 (Deployment/Fleet/Sandbox usage) · official MCP https://docs.langchain.com/langsmith/langsmith-mcp-server · npm langsmith · pip langsmith · card https://agentgid.com/apis/langsmith/agent.md
- Also fits: LangWatch, W&B Weave, Ragas, Pydantic Logfire (https://agentgid.com/apis/observability/)

## Model (all of these can call tools; prices per 1M tokens)
- Default: fast and cheap: Claude Haiku 5.5 ($0.1 in / $0.5 out) · https://agentgid.com/models/claude-haiku-5-5/agent.md · or GPT-6 Luna, MiMo V2.6 Flash
- Harder questions: Gemini 3.8 Flash ($0.75 in / $3.75 out) · https://agentgid.com/models/gemini-3-8-flash/agent.md · or Kimi K2.7 Code, DeepSeek V4 Pro
- Hardest work: Kimi K3 ($3 in / $15 out) · https://agentgid.com/models/kimi-k3/agent.md · or Claude Opus 5.5, Claude Fable 5.1

## Or start from a ready platform
- ElevenLabs Agents: free tier, from $6/mo · https://agentgid.com/agents/elevenlabs-agents/agent.md
- LiveKit Agents: free tier · https://agentgid.com/agents/livekit-agents/agent.md
- Retell AI: free tier · https://agentgid.com/agents/retell-ai/agent.md
- Vapi: free tier, from $29/mo · https://agentgid.com/agents/vapi/agent.md

## Libraries
- `pip install pipecat-ai`: Pipecat (BSD-2-Clause) · https://agentgid.com/tools/pipecat-ai/agent.md
- `npm install ai`: AI SDK (Apache-2.0) · https://agentgid.com/tools/ai/agent.md
- `pip install fastapi`: FastAPI (MIT) · https://agentgid.com/tools/fastapi/agent.md
- `pip install pydantic-ai`: Pydantic AI (MIT) · https://agentgid.com/tools/pydantic-ai/agent.md

## Agent Skills (read a skill before you install it: skills can run scripts)
- `npx skills add openai/skills --skill speech`
- `npx skills add krillinai/OpenCreator --skill krillinai-tts`
- `npx skills add cursor/plugins --skill add-voice`
- `npx skills add anbeime/skill --skill qwen3-tts-local`
- `npx skills add midudev/autoskills --skill elevenlabs-tts`
- `npx skills add MengTo/Skills --skill elevenlabs-tts`

## Coding agent to build it
- OpenAI Codex: free tier, from $8/mo · https://agentgid.com/agents/openai-codex/agent.md
- Claude Code: from $20/mo · https://agentgid.com/agents/claude-code/agent.md
- OpenCode: free tier, from $10/mo · https://agentgid.com/agents/opencode/agent.md

## Also on AgentGid
- Video and audio to text kit: https://agentgid.com/build/kits/transcribe/
- Voiceover sources: https://agentgid.com/assets/voiceover/
- Sound effects: https://agentgid.com/assets/sound-effects/

## Before you ship
- Calls carry personal data: say at the start that the caller is talking to an AI and whether the call is recorded, where the law asks for it.
- For health or other sensitive data (a clinic), check whether each service signs the agreement your rules need (a BAA under HIPAA in the US, a DPA under GDPR) before real callers use it.
- Keep API keys on the server, never in the client; set spending limits where the provider offers them.
- Read each service's terms for your use (data retention, commercial use, rate limits).
- Log model calls (traces) and test with your own real inputs before launch.

Recipe page: https://agentgid.com/build/recipes/voice-agent/ · all product recipes: https://agentgid.com/build/recipes/agent.md