Logo
SoundFlare
v0.2Free & open source · MIT licensed

The flight recorder
for voice AI agents

Open-source observability for LiveKit and Trillet agents. Trace every turn, catch hallucinations and wrong actions automatically, and run it all on your own infrastructure.

Star on GitHubSelf-host in one command
localhost:8000
SoundFlare observability view showing a call's turns, latency, and pipeline operations
LiveKitTrilletOpenAIGeminiElevenLabsGoogle TTSSarvamSupabaseDockerLiveKitTrilletOpenAIGeminiElevenLabsGoogle TTSSarvamSupabaseDocker
01 — Capabilities

See what your agent actually did.

Build, test, and debug voice agents from one dashboard you own. No per-call pricing, no seat limits.

Core

AI call reviews

A Gemini-powered reviewer reads every call and checks the agent's claims against what actually happened: tool calls, API responses, and your knowledge base. Results stream in live.

call_a41f · ai review 2 issues
callerCan you move my appointment to Friday?
toolget_availability(date="fri") 200 OK
agentDone! You're booked for Friday at 3pm.
Wrong action · reschedule_appointment was never called
agentYour copay for that visit is $20.
Hallucination · not in knowledge base
auto-review per agentbatch review recent callsqueued on every new calllive updates via SSE
Traces

Span-level waterfall

STT, LLM, tools, and TTS for every turn.

stt
llm
tool
tts
0ms500ms1s
Metrics

Latency, tokens & cost

471ms · ttft p50
tokens 1.2kcost $0.018

Automated evaluations

AI callers that behave like real people stress-test your agent on the happy paths and the messy ones.

Outbound campaigns

Upload a CSV, set call windows and retries, and watch every dial land in the logs.

Agent builder

Create and deploy LiveKit agents with your pick of LLM, STT, and voices from ElevenLabs, Google, or Sarvam.

SIP & telephony

Manage trunks, dispatch rules, and phone numbers so agents can take and place real calls.

Search & saved views

Global search, advanced filters, and saved views across thousands of calls.

Voice bug reporting

Flag issues out loud mid-test. Voice commands pin the exact moment in the call.

02 — Self-host

Up and running in one command.

A Docker Compose stack with the dashboard, Postgres, Supabase Auth, and an API gateway. Calls, recordings, and transcripts stay inside the compliance boundary you already manage, and AI reviews can run on your own cloud account or a model you host.

  1. 01

    Run the stack

    Generates JWT keys, starts every service, and seeds a login.

  2. 02

    Instrument your agent

    Add the Python SDK to your LiveKit or Trillet agent.

  3. 03

    Watch calls land

    Traces, metrics, and AI reviews show up in real time.

Docker · Vercel · any VM MIT licensed
$ git clone https://github.com/TrilletAI/Soundflare.git
$ cd Soundflare
$ ./scripts/docker-start.sh
# dashboard → http://localhost:8000
# login credentials are printed in your terminal

SDK source: TrilletAI/soundflare-sdk

03 — Open source

Star it. Fork it.
Ship it.

SoundFlare is free and built in the open. Open an issue, send a pull request, or just tell us what your voice stack needs.

Star on GitHubOpen an issue