Phone, messaging and web
Inbound and outbound calls on Twilio, Telnyx or your own Asterisk PBX, plus WhatsApp, SMS, browser calls and a Chat API. One agent answers on all of them.
v2.0.0Agent evals, tool calling on text, and pronunciations→
Open-source infrastructure for AI voice agents
TurnCall runs real-time AI agents on phone calls, WhatsApp, SMS and the web. Choose your own speech, language and voice providers, deploy on your own infrastructure, and keep full control of your call data.
- Self-hosted
- No per-minute platform fees
- Twilio, Telnyx and Asterisk
Channels
PhoneTwilio, Telnyx, Asterisk
WhatsAppVoice calls and messages
SMSTwo-way text conversations
WebWebRTC calls and Chat API
Live call on Phone00:42
- 1Speech-to-textDeepgram, AssemblyAI, Cartesia, ElevenLabs, OpenAI
- 2Language modelClaude, GPT, Gemini, Bedrock, Ollama
- 3Text-to-speechCartesia, ElevenLabs, Deepgram, OpenAI
CallerCan I move my cleaning to Friday?
AgentFriday at 9:30 is open. Shall I book it?
Turn detectionInterruptionsToolsKnowledge baseVoicemail detection
Your systems
Call routingChoose the agent for each call
Tools and MCPYour APIs, during the call
WebhooksSigned events for every call
StorageRecordings in your own S3
v2.0.0Latest release
174Commits to main in the last 30 days
MITOpen-source license
Docker · HelmRuns on your own servers
Works with the carriers and AI providers you already use
Platform
Everything a voice agent needs, in one service
Define an agent once and run it on every channel, with the models, tools and data it needs to handle a real conversation.
Any model provider
Choose each stage per agent, or use one speech-to-speech model. Run models locally with Ollama.
Tools and integrations
Call transfer, agent handoff and keypad tones built in, plus your own webhook tools and MCP servers.
Knowledge bases
Upload PDFs and documents and let agents answer from them, with hybrid vector and full-text search.
Natural conversation
Model-based turn detection, interruptions, voicemail detection and pronunciation control for brand names.
Call analysis
Summaries, sentiment, success scoring and the structured fields you define, extracted after every call and delivered to your webhook with the transcript and latency.
Performance
Low latency you can measure
TurnCall records the timing of every turn in every call, broken down by speech recognition, model and voice. You see exactly where the time goes and which setting to change.
- ~800 msProcessing time per turn with a speech-to-text, model and voice pipeline
- ~300 msProcessing time per turn with a speech-to-speech model
- p50 / p95Latency summary included with every completed call
Response time for one turn
Default settings
End-of-turn detection
Speech-to-text
Model, first token
Voice, first audio
Speech-to-speech model
Typical figures with TurnCall's default settings, measured from the moment the caller stops speaking. End-of-turn detection is adjustable per agent, and actual times depend on the providers you choose.
Developers
An API for every part of the call
Create agents, connect numbers, place calls and receive events from your own code. TurnCall has no required dashboard, so it fits into the systems you already run.
- REST API for agents, numbers, calls, campaigns and knowledge bases
- Signed webhooks for call events, with the transcript and analysis
- A call-init webhook to pick the agent and context for each call
- Live call control: transfer, hand off, send keypad tones or end the call
- Agent versions with one-step rollback and A/B traffic splits
Build with AI
Let Claude Code wire it into your app
The TurnCall plugin for Claude Code knows the API, the webhooks and call-init, so it adds calling to your app the way TurnCall expects and leaves your business logic where it belongs.
- Carries a curated API reference and a snapshot of the OpenAPI spec
- Starts on its own whenever you work with TurnCall
- Or copy skills/turncall into your project's .claude/skills
claude
> /plugin marketplace add kobikis/turncall-skill > /plugin install turncall@turncall-skill Installed turncall@turncall-skill > Call every new lead in our CRM within five minutes, using the sales agent. Example prompts. Claude uses the plugin to pick the right TurnCall endpoints and webhooks, and asks you where your own logic should go.
Production
Ready for real customer traffic
Reliability, testing and security features for teams running voice agents on live phone lines.
Provider failover
Configure backup speech, model and voice providers. If one fails during a call, the next takes over and the caller still gets an answer.
Deploys without dropped calls
When a server shuts down, live calls finish first. Readiness probes and a Helm chart are included.
Automated testing
Run scripted or simulated conversations against the real pipeline, in text or audio, and block a release in CI when an agent regresses.
Outbound campaigns
Call a list within each contact's local calling hours, with pacing to your number's rate, retries by outcome and a do-not-call list.
Usage limits
Cap live calls per project, outbound calls per second and messages per minute, enforced across every server.
Secure by default
Signed webhooks, per-call media tokens, encrypted provider credentials and project-scoped API keys with roles.
Use cases
Built for the calls your team takes every day
Reception and scheduling
Answer every call, book and move appointments, and transfer to staff when needed.
Inbound · tools · transferReminders and follow-ups
Call customer lists at the right local time, with retries for missed calls and voicemail messages.
Outbound · campaignsCustomer support
Resolve common questions from your documentation and hand off to a person with a briefing.
Knowledge base · warm transferLead qualification
Ask qualifying questions, extract structured answers and send them to your CRM after the call.
Analysis · webhooks
Self-host
Runs on your infrastructure
TurnCall is self-hosted: there is no TurnCall cloud to sign up for. It's a Python service backed by PostgreSQL and Redis that runs wherever you run containers, and you pay your providers directly.
- Recordings and transcripts stay in your own database and storage
- Each project can bring its own provider keys, encrypted at rest
- MIT-licensed source code your security team can review
Docker Compose
make docker-up && make migratePostgreSQL, Redis and the API for local development. Quickstart
Kubernetes
helm install turncall ./charts/turncallAPI, evaluation and campaign workers, with autoscaling on live calls. Kubernetes guide
Container image
ghcr.io/kobikis/turncallPublished for every release and every merge to main.
Is there a hosted version?
Not today. TurnCall is self-hosted: you run it with Docker Compose or the Helm chart, next to PostgreSQL and Redis, on infrastructure you control.
What does it cost?
TurnCall is free and MIT licensed. You pay your phone carrier and your AI providers directly, at their own rates. There is no per-minute platform fee on top.
Which phone carriers does it work with?
Twilio and Telnyx numbers, inbound and outbound, and your own Asterisk PBX over ARI (Asterisk 20.18, 22.8, 23.2 or later). WhatsApp Business voice and messages run through the WhatsApp Cloud API.
How fast does the agent respond?
Processing takes about 800 ms per turn with a speech-to-text, model and voice pipeline, and about 300 ms with a speech-to-speech model. On top of that comes end-of-turn detection, 800 ms by default and adjustable per agent. Every call records its own timings, turn by turn.
Which AI providers can I use?
Speech-to-text: Deepgram, AssemblyAI, Cartesia, ElevenLabs and OpenAI. Language models: Anthropic, OpenAI, Google Gemini, AWS Bedrock, OpenRouter, Ollama and any OpenAI-compatible endpoint. Voices: Cartesia, ElevenLabs, Deepgram and OpenAI. Speech-to-speech: OpenAI Realtime, GPT-Live-1, Gemini Live and Amazon Nova Sonic 2.
Where does my call data go?
Recordings, transcripts and call records stay in your own database and storage. Audio and text go only to the providers you configure, and you can run the language model on your own hardware with Ollama or any OpenAI-compatible server.
How do I test an agent before it takes calls?
With evals: scripted or simulated conversations that run against the real pipeline, with tools mocked. The turncall eval command exits non-zero when an agent regresses, so it can block a pull request. See the evals guide.
What's the license?
The TurnCall engine is MIT licensed. The optional Agent Builder apps are source-available under FSL-1.1-ALv2.
Start building with TurnCall
Follow the quickstart to create an agent, connect a phone number and take your first call. Planning a larger rollout? Write to hello@turncall.io.