Voice Agent Admin

Buy a phone number

Call-forwarding instructions

Point your existing business line at this agent so real inbound calls are answered by it. Set up forwarding once with your phone carrier — calls to your number then ring through to the agent's number below.

Forward calls to

—

This agent has no phone number yet. Attach one in the Channels tab (or buy one on the Phone numbers page) first.

Forward every call (unconditional)

📱 Mobile (AT&T, T-Mobile, most GSM): dial **21*<number># then call. Cancel with ##21#.
☎️ Verizon / many landlines: dial *72<number> then call. Cancel with *73.

Forward only when busy / unanswered (conditional)

⏳ No answer: **61*<number># (cancel ##61#)
📞 Busy: **67*<number># (cancel ##67#)
🚫 Unreachable: **62*<number># (cancel ##62#)

Replace <number> with the agent's number above (digits only, include the country code, e.g. +1…). Exact codes vary by carrier — if these don't take, set forwarding in your carrier's app or call their support and ask to forward to this number.

All Agents

No agents yet. Create one to get started.

Studios

Persona Studio

—

Edit this agent Its board tests

A persona is how the agent speaks — tone and register only. What it says stays in its talking points, and its rules stay in the agent prompt, where a tone experiment can't delete them.

Persona prompt

Draft

Saved

Not live

steady ×3 rounds

Tests for this persona — scored 0–10, pass mark 8.

How these tests work

Each test is what the guest said plus what a good reply must do. Scored 0–10; the pass mark is 8. Only the talker runs — no threads, no tools — so a routing bug can't fail a persona test. Save writes the row. Clicking out of a box saves it too, and Run saves first.

Write criteria about how it sounds, not what it does. This runs the talker alone — it cannot see threads or tools, so "hands off to a human" or "fires the transfer" will fail here no matter how the agent really behaves. Those belong to the board tests.

The prompt has changed since these scores were recorded — re-run to see what the new wording earns.

Running…

Bug Map

Every recorded bug, placed on the part of the machine it hit. Built from the bug fixtures — the permanent record every live failure gets before it gets fixed. Click a part to see its bugs.

Channels

Phone numbers

Twilio not connected. Add TWILIO_ACCOUNT_SID and TWILIO_AUTH_TOKEN to your .env file and restart the api container.

No numbers yet.

Buy your first number to receive calls.

Home

Dashboard

Recent calls

Last 24h, newest first.

Failing tests

Click to open the agent's editor and fix.

Observability

Calls

Has recording Test runs

No calls yet

Place a call from a web channel or run an api-channel test to see it here. Calls in progress refresh automatically every 10 seconds.

Call

j / k next/prev enter open esc close

Transcript

Who spoke, when, and the words — nothing else.

Turns

Each turn is a mini distributed trace: caller utterance → VAD/STT/LLM/TTS flame bar → agent response. Click a stage to drill into its raw event payload. Click Why? to see the post-turn supervisor reasoning.

Details

Final state at end of call.

New agent

Unsaved

Identity

How this agent appears in the picker, bundles, and your folders.

Name

Description

Public description shown when callers land on this agent.

Folder

Character

Who the agent is, and the rules it always follows. Every model call sees this.

Looking for the tunable voice?

That lives in Persona Studio — a short prompt, scored out of 10 by a test suite.

Character

Who the agent IS — personality, identity, behavior.

Business hours (optional)

Freeform string surfaced as {{businessHoursToday}} in any prompt or tool URL. Use it inside a persona to ground the agent: "If the caller asks about hours, say {{businessHoursToday}}."

Model

Foreground LLM — drives the agent's spoken response. Groq Llama runs on dedicated LPUs and is typically ~500ms faster per turn than gpt-4o-mini at comparable quality. Voice is configured per channel.

Supervisor model

Background supervisor — runs extractions, coverage, and tool triggers after each turn on live calls and in the Try drawer. Test runs use the picker in the Tests panel instead.

Engine models (speculative pipeline)

Per-thread response generators.

Picks the winning thread each turn (tiny context).

On-utterance variable extraction (pre-generation, fast).

Thinking brain — ≤10-word turn bullets behind playback.

Conversation

The scene, the facts, the goal — the per-call context.

Scenario

The situation — what's happening, who's calling, what's the context.

Knowledge Base

Facts the agent knows — company info, product details, FAQs, terms.

Goal

What the agent is trying to achieve in the conversation.

Variables

A webhook the brain calls once at session start. Response keys become {{key}} placeholders you can use in any prompt field, talking point, or tool description.

Method

URL

Headers (JSON, optional)

Body (JSON or string, ignored for GET)

Response

Discovered keys (use as {{key}})

Saved keys from previous test

Test presets

Named mock-CRM values for testing on the web channel. Pick one from the dropdown on /agent/<slug> before clicking Talk; the values become {{key}} placeholders for that call.

No presets yet — add one and pick it from the dropdown on the web channel.

Save the agent first, then configure variables.

Talking points

Things the agent must cover during the call. Each has a coverage condition and optional variable extractions. Coverage + extractions are evaluated by the supervisor in the background — they don't add latency to speech.

Ordered list

Drag to reorder. The agent is told to cover them in this order.

No talking points yet. The agent will rely on the goal alone.

Save the agent first, then add talking points.

Channels

How people reach this agent.

Attached channels

Web links and phone numbers connected to this agent.

No channels yet — pick how people reach this agent.

Save the agent first, then add channels.

Channels (web link, phone, API test) attach after the agent has a slug.

Tools

Functions the agent can invoke mid-call. Each tool runs as a webhook (or a built-in preset).

Attached tools

The LLM picks tools by name + description.

No tools attached yet.

Save the agent first, then add tools.

Tools (presets + custom webhooks) attach after the agent has a slug.

Tests

Calls you script, replayed against this agent, with checks that must hold. Run them before a change reaches a real caller.

Judged by Run as

⚠

Running…0.0s

Group Sort

No tests yet

Turn a real call into a test from the Inspector, or write one from scratch: what the caller says, then the checks the agent must pass.

Save the agent first, then add regression tests.

Describe what should happen

One or two sentences in plain English. The agent’s own pathway drafts the caller lines and the checks; you read them over and fix anything that’s off before running.

Try:

The draft is saved as a test and opened for review. Nothing runs until you do.

New test

Name

Mode

Flow (groups the list)

Description

Scenario

Snapshot the state from a session test (exact) or draft one from the description — then tweak below.

Agent actually said open source call ↗

What should it have said? — plain English; saved as an LLM-judge expectation that also sees the caller line

Caller utterance — the one line said at this point

Standing at — the node the agent is on

Variables gathered so far

Already covered — nodes done before this turn (picking a focus auto-checks everything before it)

Expectations

Each row is one assertion. The test passes when every row passes.

Campaigns

Outbound batches this agent dials on its own: leads pulled from the CRM, paced inside each guest's local window under a daily cap, dispositioned back when the call ends. Dry-run until every rail says live.

dry-run server

No campaigns yet. A campaign is a lead source plus a calling window, a daily cap and a pace — create one and run a dry batch first.

Batch of

Settings

Name Mode Lead source Window start (callee local) Window end Daily cap per phone Max in flight Per minute

Jobs (America/New_York)

Preview

Record	Phone	Local time	Zone	Verdict	Why

Attempts

When	Record	Phone	Status	Outcome	Disposition	Call

Settings

Limits and integrations.

Timeout (seconds)

Max conversation duration. 0 = no limit. Default 300 (5 min).

Interruption sensitivity

Min sustained speech (ms) before a caller can barge in. Higher = harder to interrupt.

Min words heard.

Blank = server defaults. Applies to live calls; the opening greeting is never interruptible.

Emotion

Colours generated replies with the agent's current emotional state, and lets it earn the occasional audio tag.

Expressiveness 0-10. 0 = never uses a tag, 10 = almost every turn.

Verbatim scripted lines are spoken exactly as written and are never affected.

Webhook URL

Posted-to when a call ends with the full transcript + metadata.

Phone forwarding

How to forward your existing business line to this agent's phone number.

Variables