Every conversation, fully observable.
One timeline shows transcript, audio replay, sentiment, tool calls, and the latency breakdown across STT → LLM → TTS. Filter by agent, vertical, language, sentiment, or escalation. Export to S3 or your SIEM in one click.
Production agents fail in ways logs never explain.
A voice agent does not just have a stack trace. It has a turn that took too long, a tool call that returned the wrong shape, a sentiment that cratered at the 90-second mark, and a TTS provider that introduced 600ms of jitter on a Tuesday afternoon. The Session Viewer is the place where all five of those facts live on the same timeline.
Every Saaya session — voice, chat, video — is captured end to end. Audio is stored, transcribed, and aligned with sentiment per turn. Tool calls show their request, response, and retry chain. The latency breakdown splits the response into STT, LLM, TTS, network, and Saaya orchestration so you can see where the milliseconds went.
Filter, search, and export. Find every session where the agent escalated to a human in the last 30 days, where sentiment dropped below 0.3, or where a specific tool call timed out. Export the matching sessions — full audio, transcript, metadata — to S3, Snowflake, or your SIEM for compliance review.
The observability layer your agents always needed.
Built for the support engineer chasing a regression and the compliance officer who needs an audit trail.
Per-session replay
Synced audio, transcript, sentiment, and tool calls — scrubbed on a single timeline. Click any turn to jump the audio cursor. Compare two sessions side by side.
Latency breakdown
Per-turn split across STT → LLM → TTS → network → Saaya orchestration. Tail latencies and outliers flagged automatically. Find the leg dragging your p95.
Sentiment band
Per-turn sentiment from -1 to +1, smoothed across the conversation. See the moment a call went sideways and the exact prompt that turned it around.
Tool-call inspector
Every function call with the full request, response, retry chain, and timing. Spot mis-shaped responses, slow integrations, and retry storms in seconds.
Transcript search
Full-text search across millions of sessions. Filter by agent, vertical, language, sentiment threshold, escalation flag, or arbitrary metadata you tagged at runtime.
Audit-log export
One-click export of any filter set to S3, Snowflake, BigQuery, or a SIEM (Splunk, Datadog, Elastic). Includes audio, transcript, metadata, and structured tool-call logs.
Three steps from session to insight.
Capture is automatic. The work is figuring out what to look at.
Capture every session
Every voice, chat, and video session is captured end to end the moment it lands on a Saaya agent. Audio, transcript, sentiment, tool calls, and per-leg latency — no instrumentation required.
Open the timeline
Pick a session from the dashboard or land on it from a webhook. The unified timeline shows every signal aligned to the second. Scrub, jump, compare.
Replay, search, export
Replay live, search across millions of sessions, save filters as views, export matching sets to S3 or your SIEM. Wire up alerts on sentiment drops, escalations, or latency outliers.
Sessions are first-class API objects.
List, fetch, and export sessions like any other Saaya resource. Cursor-paginated, shape-stable, audit-ready.
# List sessions — filter by agent, sentiment, escalation
curl "https://api.saaya.ai/v1/sessions?\
agent_id=agent_renewals_v3&\
sentiment_lt=0.3&\
escalated=true&\
since=2026-04-01&\
limit=50" \
-H "Authorization: Bearer $SAAYA_API_KEY"
# {
# "data": [
# { "id": "ses_01HXAB...", "agent_id": "agent_renewals_v3",
# "started_at": "2026-04-12T10:14:22Z", "duration_ms": 184220,
# "sentiment_min": 0.18, "escalated": true,
# "latency_p95_ms": { "stt": 240, "llm": 880, "tts": 320 } }
# ],
# "next_cursor": "cur_abc..."
# }On every tier — what you get.
Session capture is included on every plan. The difference is retention and export depth.
7-day session retention · transcript + sentiment · basic latency view · no audio replay · no export.
90-day retention · audio replay · full latency breakdown · tool-call inspector · transcript search · CSV export · saved filter views.
7-year retention (or BYO retention policy) · S3/Snowflake/BigQuery/SIEM destinations · BYO storage · region-pinned recordings · DPA · 99.95% SLA on retrieval.
Pair this with the right Solution.
Frequently Asked Questions.
Yes by default — voice, chat, and video. You can disable audio capture per agent (transcripts still capture) or per session at runtime via consent flags. Per-region defaults follow local consent law (two-party in CA, GDPR in EU, DPDP in IN).
Watch the calls.
The single fastest way to make your agents better is to actually see them in production. Capture is on the moment your first session lands.