Skip to content

After the Call

The call path — records live in the highlighted hop

Everything in this chapter lives in the highlighted hop: Parrot Cloud keeps the transcript, recording, and report of every call.

Every call — answered, missed, or failed — lands in Call history with a transcript, a verdict on what happened, and (a moment later) an AI-written recap. This page shows you where to find each piece and how to read it.

The Call history page

Open Call history in the console. Every call across your devices is listed newest-first, searchable by caller, assistant, or title, and exportable to CSV. In the All / Answered / Missed filter, Missed means everything that didn't become a two-way conversation — rung-out calls, failures, and calls your assistant gave up on alike.

Call history: outcome chips, duration, credits per call

Each row's headline is the AI-written title of the call once it's ready; until then it shows the assistant's name. The row also pins the exact assistant version the call ran (the @v… tag — click it to view that version read-only). Click the row to open the transcript.

The list loads the most recent 200 calls; Load more extends it up to 500, the most the console shows. Search and the CSV export cover what's loaded.

The call kinds

Under each title, the row says who took the call live — the same fact the board itself attributes, so History can tell an assistant call from your own at a glance:

Shown as Who took it What the row means
Parrot answered The board picked up the ring, or you/your agent answered it from the console An assistant call, incoming (↙)
Parrot dialed Placed from the console or by your agent An assistant call, outgoing (↗)
Your call You answered or dialed on the phone Your conversation — the assistant was on it only while you routed the call's audio to the board (Who's on the call)

The kind says who started it; how much of the call the assistant was actually on is its own fact, because the assistant can leave and rejoin any call mid-way (the phone's audio-route selector). A row where the assistant was on for only part of the call shows a small "Parrot on m:ss" chip; a call it never joined reads dimmed with "Parrot not on this call" — a timestamp for your line's records, with no transcript, no recording, and no charge. Calls the assistant carried end to end — the normal case — show nothing extra.

Older calls (before this attribution existed) and calls that only rang show no kind — nothing is guessed after the fact.

Transcript & replay

The transcript view shows the conversation turn by turn — the other party and your assistant — synced to the actual recorded audio. Click any line to play from that moment; the rail has a scrubber, per-side level meters and mutes.

Transcript view: summary card, turn-by-turn conversation, recording player

Once the AI summary is ready, it appears as a card above the transcript with a Goal met / Goal not met chip — click it to open the full call report. If there's no AI summary, a Report card still opens the full report. The card also tells you when call statistics are being prepared or are unavailable; a missing statistics result does not remove the AI summary.

Delete call history

For one finished call, use Delete call history at the bottom of the replay details or the report. To delete several, use Select calls in Call history and choose the finished calls to remove. Review the confirmation before deleting. An MCP agent can use delete_calls(call_ids) when you ask it to delete specific calls. Its key can delete only your finished calls on its bound device, even though list_calls can read history across your account's devices.

The call disappears from history immediately, and its old replay/report links stop working. Conversation content, recordings and saved files are cleaned up separately. Credits already used, billing transactions and usage totals stay unchanged. You cannot restore the call from the console.

Downloaded copies stay on your device.

The call report

The report is the full picture of one call, split into clearly-labeled halves so you always know what's measured and what's AI-written. The header line carries the direction, the number, the assistant @version and the credits used; the footer names the device, the phone, the firmware and bridge versions, and the call ID. Top to bottom:

  • Summary — a short recap (what happened, did it succeed, any follow-up needed), explicitly labeled as written by AI.
  • Call health — Normal, Degraded, Failed or Unknown, with anything that went wrong on the line during the call (audio dropouts, a connection restart…), most serious first — or "No confirmed call failures detected."
  • Local audio delivery — how much of the phone's audio reached Parrot, and how much of the AI's audio was accepted for playback (what the other side actually heard is not measured).
  • Call timeline — what happened when, counted from the start of the call.
  • Extracted data — if your assistant defines output variables (a confirmation number, a quoted price — see Customizing your assistant), whatever the call produced appears here. A field marked req in the editor is one every call should capture; when it is missing, the report shows not captured. Goal met is the AI's own judgment of the call's objective and does not check these fields.
  • Call statistics & technical details — measured, not AI-generated: duration, how long the AI session was active (and how much of the call was outside it), how quickly the call connected and the AI first responded, and transport/audio-quality numbers.
  • Link & device measurements and Buffer statistics — the Bluetooth, bridge and buffer counters, mostly for support conversations.

The JSON button at the top shows the raw numbers for the whole report.

The call report: AI summary up top, measured statistics below

Two different \"did it work?\" answers

The report's outcome answers "did the call connect and reach a person?" The summary's Goal met flag answers "did the AI accomplish what you asked?" A call can connect fine and still fail its goal — or vice versa. The report never merges the two.

Is the AI's summary trustworthy?

The summary is labeled as AI-written and judged against what you asked the call to do. The measured parts of the report — duration, outcome, who hung up, audio quality — are never AI-generated, so you can always fall back to the transcript and the numbers.

What "outcome" means

A finished call's outcome is determined from what happened (who spoke, who hung up, what failed), rather than an AI assessment:

Outcome Shown as Meaning Worth redialing?
answered Answered A real two-way conversation happened. No — it connected.
no_answer No answer Nobody engaged — it rang out or went nowhere. Yes
agent_hangup Agent hung up Nobody engaged and your assistant chose to hang up. No
needs_human Needs human Your assistant escalated — it needs your input. No — it needs you
error Failed Something failed on our side or the AI vendor's. Sometimes — transient failures yes

Alongside it, ended_by records who actually closed the line: your assistant, the other party, a dropped connection, an explicit stop from you, or a server fault. The outcome is available the moment the call ends; the summary and extracted data can take a little longer, and the outcome never waits for them. Some older call records may have no outcome; see Call outcomes and retry rules.

Troubleshooting

There's no summary on my call

The summary is written a few moments after the call ends — refresh. A call with no conversation (nobody answered) gets no summary at all; the outcome and statistics are still there.

The report has no Extracted-data section / fields weren't captured

The section appears when the call produced something to show. The assistant only fills a field if the conversation actually yielded it — check the transcript; if the information was said but not captured, give the field a clearer name under the editor's Tools → Output variables (confirmation_number rather than number), or say in the assistant's prompt what it must ask for (Customizing your assistant).

Where this fits

Reading calls is half the loop — the other half is making the assistant better at them. That's next: Customizing your assistant.