Dograh
Monitoring and QA

Find the bad call, then read its trace

Dograh is the self-hostable voice agent platform with an open-source workflow builder. Dograh reviews each finished call and flags the ones that went wrong, and the trace behind a flag shows which part of the agent caused it.

6,200+GitHub starsBSD-2open source$0platform fee, self-hostedAny systemconnects to the tools you use
01 / How it works

Review, trace, then fix

Example
QA flags DEAD_AIR on a live callthe trace shows a correct transcriptso the prompt is at faultreplay that step
01 · Review

Grade calls after they end

QA nodean AI review of the transcript

It runs once the call is over, so it never changes what the agent says. Each step gets its own tags, a 1-10 score and a short summary.

→ Swap the default prompt for your own checks

02 · Trace

Open the call behind a flag

View Traceevery layer of one call

See what speech-to-text heard, the full prompt and each tool request with its response. A wrong transcript is a speech problem no prompt edit fixes.

→ It also shows speech, AI and voice time per turn

03 · Replay

Repeat the failing step

Langfuseone AI step, rerun with its exact prompt

Run that step several times before you touch the prompt. A step that fails once in five runs will fail on real callers too.

→ Keep Langfuse traces private

02 / What it works with

Built in, plugged in, and where it lives

Built into Dograh
QA nodeView TraceCall TranscriptRun DetailWebhook annotations

QA results show in Run Detail and reach your webhooks too

Tools that plug in
LangfuseTunerNoveumBigQuery

Langfuse keys are set per organisation, on managed or self-hosted Dograh

Stays on your servers with
Self-hosted DograhSelf-hosted LangfuseOpen-weight models you run

A hosted model provider on your own keys still receives the call audio

03 / Sample rate

How many calls to review, and when

  1. 100% at launchSample Rate (%) starts at 100. Leave it there for the first weeks live.
  2. 100% after any flow changeGive every change a week of full review before you lower it again.
  3. 5-10% once stableEach review is a model call, and someone has to read every flag.
  4. Regulated scripts at 100%Where a missed disclosure is a legal problem, keep reviewing every call.
  5. Phone path testedMost phone calls still reach the agent as 8 kHz audio, while a browser test runs at 16 kHz.
04 / Why Dograh

Monitoring you can run yourself

Measures without interfering

The QA node has no edges in the flow, so live calls run as you built them.

A score for every step

Tags and scores sit on each node, pointing straight at the step that slipped.

Traces on your servers

Self-host Dograh and Langfuse and run the models yourself, the QA model included, and nothing leaves your network.

No platform fee self-hosted

Self-hosted Dograh charges no per-minute fee. Servers and any paid models still cost money.

05 / Guides

Learn how it works, step by step

Complete guideHow to Test, Monitor, and Scale AI Voice Agents in Production: Start With a Phone CallRead →TutorialHow to Load Test a Voice Agent Through the Real Phone PathLoad test a voice agent through your real carrier and number: what a browser test skips, how to size the test, three ways to place calls, what to record at each step and how calls end at the concurrency cap.TechnicalChanging What Your Voice Agent Says During an OutageHow to change what a voice agent says when a system it depends on is down: a notice switch in your own system, a recorded outage agent, fallback lines for slow tools, pausing campaigns and switching back.TechnicalGetting Names, Numbers and Emails Right in Voice AgentsWhy voice agents mishear spelled names, numbers and emails on phone calls, what public voice tests measured, and the capture loop that fixes it: spelling, read-back, validation and a handover.Industry InsightWhy Most Voice Agent Demos Break on Real Phone CallsWhy voice agent demos break on real phone calls: 8 kHz phone audio, callers who pause and interrupt, voicemail, transfers and slow lookups, and how Test Chat, real phone tests and call traces find each one.TechnicalHow AI Voice Agents Handle Accents and Background NoiseWhy accents and background noise still break voice agents, and the fixes in order: speech-to-text chosen on your own callers' audio, a short dictionary, interruption settings that ignore noise, prompt rules for misheard words, and testing on real phone calls.TechnicalMultilingual Voice Agents: Serving Callers in Their Own LanguageHow a multilingual voice agent detects and switches language mid-call, which speech-to-text options switch and which stay fixed, why one voice must speak every language, what the agent should do when it is unsure, and how to test each language.TutorialTesting Voice Agents Before You Go Live: Five Tests, in OrderFive tests to run before a voice agent takes real calls: a web call, simulated callers, repeat runs per leg, a real phone call and a small load test.TechnicalAutomated QA for Voice Agents: Grade the Agent, Not the CallerWhat Dograh's QA node checks after each call, and why a 5-10% sample of live calls is enough to find common failures within days.TechnicalPrompt and Flow Design Patterns for Reliable Voice Agents: Split by GuaranteeWhere to split a voice agent's call into nodes, where exact wording comes from, and how to prove each leg holds up over repeated runs.TutorialHow to Read Call Traces to Debug a Voice Agent: Find the LayerOpen one bad call's trace, find the first layer that went wrong, and replay that step several times before you change the prompt.
Questions
Does QA change what the agent says on a live call?+

No. The QA node runs after the call ends, on the transcript. It has no edges in the workflow, so it measures the call without changing it.

What does the default QA review return?+

For each step of the call: tags such as DEAD_AIR and USER_FRUSTRATED, a 1-10 score, a sentiment and a short summary. You can swap in your own checks.

Is the sentiment label the customer's satisfaction?+

No. It is the review's reading of how the call went, useful for sorting calls. To learn how customers felt, ask them and record the answer.

How many calls should QA review?+

Every call at launch and for a week after any flow change. Once the agent is stable, set Sample Rate (%) to 5-10%.

Can I send traces to a self-hosted Langfuse?+

Yes. Enter your Langfuse host and keys in Platform Settings for your organisation. It works on managed and self-hosted Dograh, and traces are private by default.

What happens if I make a trace public in Langfuse?+

A public trace exposes the full call transcript, the prompts and the tool payloads. Keep traces private unless you mean to share one.

Can I see how long each stage of a turn took?+

Yes. View Trace shows speech-to-text, model and voice time on each turn. To analyse many calls at once, export them to Noveum or BigQuery.

Does any call data leave my servers?+

Not if you self-host Dograh and Langfuse and run every model yourself, the QA model included. A hosted model provider gets the audio, even on your own keys.

Find the bad calls. Keep every trace.

Self-host in one command, or start on Dograh Cloud.

$ curl … start_docker.sh && ./start_docker.sh