Question 1

How does CallingBox actually improve my agent over time?

Accepted Answer

Every production call is automatically labeled for failure modes: hallucinations, weak handoffs, missed objections, off-policy answers, structured-output errors. Those labels get clustered into ranked failure types, each one paired with a proposed prompt or policy diff. You review the queue, ship the diffs you like, and CallingBox runs them through a counterfactual A/B harness before they reach 100% of traffic. The result is a weekly lift report showing which fixes moved your outcome metric and by how much.

Question 2

What's a 'ranked diff' and do I have to ship it?

Accepted Answer

A ranked diff is a concrete, reviewable change to your agent (usually a prompt edit, a tool-use policy tweak, or a structured-output schema fix) that we believe will lift the metric you care about. Every diff comes with the failure modes it addresses, the calls it would have changed, and an expected uplift estimate. Nothing ships automatically: review and approval are human-in-the-loop by default. Once approved, a diff still has to clear the A/B harness before it goes to 100% of calls.

Question 3

How do I know it's actually getting better?

Accepted Answer

Outcome-denominated reporting. You define what success looks like: dollars collected, appointments booked, verifications passed, dispositions resolved. We attribute lift to specific shipped diffs through counterfactual A/B testing on real production traffic. You get a weekly lift email, an in-app dashboard with per-diff impact, and a queryable history of every change since launch.

Question 4

Will callers know it's AI?

Accepted Answer

Most won't, but we don't recommend pretending. CallingBox uses sub-500ms turn latency and natural-sounding TTS voices, so the conversation feels human. We strongly recommend disclosing AI usage where required by law (e.g., TCPA, state-level AI disclosure rules), and the agent can be configured to introduce itself as an assistant on connect.

Question 5

What happens if the AI gets stuck or fails mid-call?

Accepted Answer

Every call has guardrails: turn timeouts, hallucination detection, and an escalation path. If the agent can't make progress, it can hand off to a human number you specify, leave a voicemail with the structured context, or end the call cleanly with a failure reason in the JSON returns. You only pay for connected time, and webhooks fire on every terminal state so your workflow always knows what happened. Every stuck-or-failed call also feeds the improvement loop: those failures become candidate diffs faster than the working calls do.

Question 6

Is this safe for healthcare, finance, or collections use?

Accepted Answer

Calls are encrypted in transit, and we don't train on your call content. We're working toward SOC 2 Type II and offer BAA / data processing agreements on the Enterprise plan for HIPAA-adjacent workloads. For regulated use cases (collections (FDCPA), healthcare (HIPAA), eligibility/verification), the improvement loop is compliance-aware: shipped diffs are checked against your policy library before they roll out. Contact us before launch in any regulated vertical.

Question 7

How is voice quality compared to other platforms?

Accepted Answer

Voice quality is week-1 table stakes: sub-500ms latency, natural-sounding TTS, low WER on accented speech. The interesting story is what happens after week one: most runtimes plateau, and ours keeps moving the metric you actually care about. Quality is the entry point; improvement is the moat.

Question 8

What does CallingBox actually do?

Accepted Answer

You configure a reusable agent with a persona, instructions, tools, and a structured returns schema. Dispatch outbound calls via POST /v1/calls by referencing the agent, or attach the agent to a number and let it answer inbound calls. CallingBox handles telephony, voice models, and the conversation; you get structured JSON back. Then every call feeds the improvement loop: labeled failure modes, ranked diffs, verified uplift.

Question 9

Can the AI handle inbound calls too?

Accepted Answer

Yes. Create an inbound agent, attach a phone number, and CallingBox answers every incoming call on that number using the agent's configuration. The same structured returns, transcript, and improvement-loop pipeline applies. Inbound agents are available on all plans.

Question 10

How do I get the results back?

Accepted Answer

Two ways. Poll the call status with GET /v1/calls/{id}, or receive a webhook when the call completes. Both return the structured JSON matching the schema you defined in the returns field.

Question 11

Can the AI call functions during the conversation?

Accepted Answer

Yes. You can define functions that the AI can invoke mid-call to check availability, look up an account, or book an appointment. Works with MCP servers on all plans.

Question 12

What happens if nobody picks up?

Accepted Answer

CallingBox detects voicemail, busy signals, and no-answer. The call status updates accordingly. You only pay for connected time.

Question 13

Can CallingBox navigate IVR menus?

Accepted Answer

Yes. CallingBox detects and handles IVR prompts, DTMF inputs, and automated phone trees. Your agent can navigate hold menus, press extensions, and reach the right person or department before starting the conversation.

Question 14

What languages are supported?

Accepted Answer

English is fully supported. Spanish is in beta. We're adding more languages. Contact us if you need a specific one.

Question 15

Which countries are supported?

Accepted Answer

Currently the United States only. We support local US phone numbers for both outbound and inbound calls. International support is on the roadmap.

Question 16

Will my calls get flagged as spam?

Accepted Answer

We sign every call with STIR/SHAKEN A-level attestation, manage number reputation, and enforce calling-hour regulations. On the Enterprise plan, we also handle do-not-call registry scrubbing and consent management. Your calls show up as verified, not as spam.

Question 17

What can I configure on each call?

Accepted Answer

Persona, voice, language, tools, returns schema, and voicemail behavior all live on the agent. When you dispatch an outbound call, you can replace the agent's default instructions for that single call, pass per-call context, and pick any caller ID already attached to the agent. All via the API, no dashboard required.

Question 18

What's included in the per-minute price?

Accepted Answer

Everything. Telephony, speech-to-text, language model, text-to-speech, orchestration, voicemail detection, structured data extraction, and the improvement loop. There are no separate line items for each component, and no add-on for the optimization layer.

Question 19

How does billing work?

Accepted Answer

You start with $5 in free credits. After that, you pay per minute of connected call time at $0.05/min. There are no monthly minimums, no platform fees, and no per-component charges. You can see your usage and remaining balance in the dashboard at any time.

Question 20

Can I cancel anytime?

Accepted Answer

Yes. There are no contracts and no cancellation fees. Your account is usage-based, so if you stop making calls, you stop paying. Enterprise agreements may have separate terms.

Question 21

How does my agent invoke CallingBox?

Accepted Answer

Three canonical surfaces: the hosted MCP server (point any MCP client — Claude Desktop, Cursor, ChatGPT, OpenClaw — at https://mcp.callingbox.io/mcp), the CLI (pip install 'callingbox[cli]', then `callingbox call ...`), or the Python and TypeScript SDKs. Every surface is a thin layer over the same /v1 HTTP API, so swapping interfaces never changes behavior.

Question 22

Is there an MCP server for CallingBox?

Accepted Answer

Yes. Point any MCP-compatible agent (Claude Desktop, Cursor, OpenClaw, custom MCP clients) at https://mcp.callingbox.io/mcp with your API key as a Bearer token. The MCP server exposes tools for placing calls, fetching results, listing numbers, and creating agents — no SDK install required.

Question 23

Can my AI agent receive calls too, not just place them?

Accepted Answer

Yes. Attach a phone number to an inbound agent and CallingBox answers every incoming call using the agent configuration you defined. Each call returns the same structured JSON your outbound calls do, so the rest of your agent loop doesn't have to care which direction the call came from.

Question 24

Does my agent get the structured data mid-call or only at the end?

Accepted Answer

Both. Your agent can poll the call status via GET /v1/calls/{id} or subscribe to webhooks for real-time events (call.answered, function.invoked, call.ended). The full structured returns object is available the moment the call ends, and partial transcripts stream during the call if you need them.

Question 25

Do I need to bring my own phone numbers?

Accepted Answer

No. You can buy local US numbers directly from CallingBox in seconds and attach them to inbound agents or use them as caller ID for outbound calls. If you already own numbers elsewhere, porting is supported on the Enterprise plan.

Voice agents that get better every day

Configure an agent. Dispatch the call.

Improve your agents with every production call.

Captured

Diagnosed

Shipped

Verified

Watch a call become an improvement.

Other voice runtimes ship and stop. We ship and compound.

And yes: we're still the runtime.

See your first improvement in week one.

Common questions