Repository navigation
Releases
tring 0.4.1
Onboarding release.
tringCLI:tring studio,tring eval,tring version— one installed command instead of module paths.- Docker onboarding:
docker compose up→ Studio at localhost:8900, provider keys passed by name, sessions on a named volume. Image built and smoke-tested. - docs/DEPLOYMENT.md: three tiers (laptop, container, production calls) with honest operational notes.
- README's Studio section now covers the v0.4 features; roadmap updated.
tring 0.4.0
The platform release: tring goes from voice-agent library to the open-source platform you build voice agents in.
Studio
- Flow builder: draw conversation flows on a canvas (say / ask / branch / tool / handoff / end), compile them server-side into an agent with numbered, transition-explicit instructions. Compiles are idempotent and the graph round-trips losslessly in the spec's metadata. Precise, node-addressed errors.
- Session replay: every studio session persists as an event stream; scrub through past calls in the same panels used live, costs and tool cards included.
- Microphone testing: hold-to-talk from the browser (AudioWorklet, 16 kHz uplink), with honest degraded states when a mic or STT slot is unavailable.
Core
- Eval harness:
python -m tring.eval evals/runs YAML conversation tests (expected tools, arguments, replies, latency) plus optional LLM judges; CI-ready exit codes. - Outbound campaigns: concurrency-bounded dialing with retries, quiet hours, Twilio answering-machine detection, and a live cost-per-outcome ladder computed from real campaign counts.
- Observability: OpenTelemetry span export, per-turn latency waterfalls, and a prompt-cache regression alarm.
- Telephony: FreeSWITCH dialplan generator for generic SIP, DTMF events, transfer-to-human tool.
- S2S: Gemini Live adapter; hybrid v2 downgrades mid-call from S2S to cascade with context carried over.
- Knowledge: retrieval providers (local BM25, Chroma, Qdrant) with speculative prefetch that searches while the caller is still talking.
505 tests. Live-verified: 10/10 vendor probes, and the flow builder + replay driven end to end in a real browser.
tring 0.3.0
The friendliness release: real turn-taking, real telephony, and Tring Studio.
- VAD turn-taking: adaptive energy VAD (zero deps) and Silero ONNX, driving utterance endpointing and true barge-in: a caller talking over the bot cancels generation and synthesis mid-flight
- Streaming local STT:
faster_whisper_streamingwith incremental partials - Telephony: Twilio Media Streams transport with a stdlib G.711 codec; Twilio's played-audio marks make interruption reconciliation exact on real phone calls. Generic SIP guide in docs/TELEPHONY.md
- Tring Studio:
pip install "tring[transports]"thenpython -m tring.studio: a browser agent designer (form + YAML), live test console, cost panel with the estimated-fraction honesty chip, and an event timeline. React app ships compiled inside the wheel: no Node needed - Declared tools get studio mock handlers so you can test conversation logic with zero backend code
311 tests. All live-verified: 10/10 vendor probes including full cascade E2E.
tring 0.2.0
Provider expansion: 18 registered providers across STT, TTS, LLM, and speech-to-speech.
- STT: AssemblyAI (realtime websocket), OpenAI (whisper-1 / gpt-4o-transcribe, with a base_url seam that reaches Groq's compatible endpoint), Sarvam (saaras, Indian languages)
- TTS: Cartesia, OpenAI (with built-in 24k to 16k resampling), Sarvam (bulbul), Piper (local, $0)
- LLM: native Anthropic (cached-token passthrough) and Gemini, joining openai_compatible which reaches Groq, Cerebras, Mistral, Together, Fireworks, DeepSeek, xAI, OpenRouter, vLLM, LM Studio (see docs/PROVIDER_ALIASES.md)
- S2S: OpenAI Realtime, Ultravox
Every endpoint, auth scheme, and price was verified against official vendor docs (rate card entries stamped as_of 2026-09). New reference docs: docs/PROVIDERS.md and docs/PROVIDER_ALIASES.md.
196 tests, ruff clean, mypy --strict clean.
tring 0.1.0
First release.
Tring is an open-source voice agent stack: one agent contract, any runtime (cascade / speech-to-speech / hybrid), local-first providers, honest cost accounting.
Highlights:
- Four production latency primitives: speak-while-thinking streaming parser, tool-call choreography (dead air is a state the schema does not allow), interruption playback reconciliation, cache-safe language lock
- Cost engine with per-unit estimated flags and denominator-ladder reporting (cost per dial / connected / conversation / outcome)
- Local $0-per-minute pipeline: faster-whisper, Ollama, Kokoro
- Cloud provider starters: Deepgram, ElevenLabs, any OpenAI-compatible LLM
- Console and websocket transports, runnable examples
- 131 tests, ruff clean, mypy --strict clean
pip install tring