Duckbill
30M+ real-world actions completed

The real world doesn't have an API. We built one.

APIs for phone calls, payments, errands, and services. Your agent pings one. The task gets done, cleanly, in a way you'd put your name on.

Start with phone calls

REST API and MCP. Bring your own agent, or use ours.

Your agent

My flight was delayed 4 hours. Get me a refund.

Found booking: UA 2847, SFO to JFK
Delay confirmed: 4h 12m
Checking DOT refund eligibility...
Eligible. United handles this by phone only.
Needs a phone call to finish
POST /v1/calls · calling United Airlines…

United approved a $847 refund. It’s back on your card in 5 to 7 days.

Four capabilities. One API.

Start with phone calls. Your users will ask for the rest.

POST /v1/callsLive

Phone calls

Book it, cancel it, dispute it, confirm it. Your agent sends the goal. A person makes the call, AI only where it reliably finishes. The outcome comes back as data.

  • Get the refund United owes for the 4-hour delay.
  • Book a dermatologist who takes Aetna, weekday afternoons.
  • Find out why the claim was denied. Then appeal it.
See phone calls
POST /v1/paymentsEarly access

Payments

For every payment agentic checkout can’t reach: the vendor who only takes checks, the card read over the phone, the site that blocks bots.

  • Pay the caterer’s $1,200 deposit by phone.
  • Mail the check the contractor insists on.
  • Finish the order that blocked your agent at checkout.
See payments
POST /v1/errandsEarly access

Errands & services

Someone has to show up. Pickups, drop-offs, returns, waiting for the plumber, and the services your users will ask for next.

  • Pick up the dry cleaning before they close at 6.
  • Return the couch. Yes, the whole couch.
  • Be there when the plumber shows up between 2 and 6.
See errands & services
POST /v1/tasksLive

Concierge

Haven’t built an agent? Use ours. Your users ask for anything. Duckbill plans it, runs it end to end, and reports back under your brand.

  • We’re moving to Austin in six weeks. Handle all of it.
  • Plan the birthday: 15 kids, nut allergy, under $400.
  • Get me a plumber this century.
See concierge

Building the Ground Truth of the Real World

Every task completed generates proprietary data no one else has. Each decision point teaches our models how the physical world actually works.

Try an example below ↓

or
NowFuture
Duckbill Reliability
—
across 0 decision points
Duckbill: AI + Human
● AI: —● Human: — (0 steps)
Without orchestration
—
01Triage
Intake QuestionsTask DecompositionSequencingBrief Generation
02Market Making
Time EstimationSkill MatchingSurge PricingDynamic Pricing
03Routing
Agent SelectionTiming OptimizationTime ZoneLanguageTone CalibrationInstructionsLive Q&APII Guard
04Execution
Phone CallsEmailsPayments⬥Offline Research⬥Fact CheckingProblem Solving⬥Pick Up⬥Drop Off⬥CompileSynthesizeWait in Line⬥Photo⬥Complete Form⬥Negotiate⬥
05Quality
Live MonitoringBaseline ComparisonFailure TriageRecovery⬥Learning LoopEscalation
Triage
Market Making
Routing
Execution
Quality
Human involvement

Reality has a surprising amount of detail

Your agent isn't unreliable. The real world is.

63%

AI alone fails 63% of complex real-world tasks.

93%

With Duckbill, the same task types resolve 93% of the time.

Three layers. AI owns one.

Every real-world task has three layers. Models have mastered the first. Duckbill runs the other two.

LAYER 3Making It Happen
LAYER 2Real-World Intelligence
LAYER 1The Plan
Layer 3DUCKBILL

Making It Happen

Phone calls, payments, pickups, negotiations, follow-ups. 33M businesses in the US have no API. This is where every AI product breaks, and where Duckbill operates.

Layer 2DUCKBILL

Real-World Intelligence

Who picks up at 9am vs 2pm? Which line gets you a person? What triggers a callback? None of this can be scraped or pre-trained. It’s learned from 30M+ real interactions, and it works before the first dial tone.

Layer 1AI

The Plan

Understanding the goal, identifying the steps, sequencing the actions. This is reasoning, and AI is excellent at it.

You ping the API. The task gets done.

One request in. Status while it runs. A verified outcome back. Everything in between is our problem, not yours.

duckbill · phone call

Request

POST /v1/calls{  "goal": "Book a new-patient appointment. Aetna PPO. Weekday afternoons.",  "to": { "name": "Dr. Patel Dermatology", "phone": "+1 617 555 0142" },  "on_behalf_of": "Jane S.",  "callback_url": "https://yourapp.com/hooks/duckbill"}

Result · webhook, a few minutes later

{  "status": "completed",  "outcome": "booked",  "summary": "Thursday Oct 2 at 2:30pm with Dr. Patel. Bring insurance card and photo ID.",  "confirmation": "APT-88213",  "attempts": 1}

REST API with webhooks, or the Duckbill MCP server for Claude, ChatGPT, and agent frameworks. MCP docs

Different tasks. Same building blocks.

What it actually takes to get things done. Every step is a place the real world can say no.

OFFLINE RESEARCHOFFLINE TRANSACTIONDIGITAL TASKTROUBLESHOOTINGPHYSICAL ERRAND

Pick up dry cleaning

Find hours & policies
Call to confirm order ready
Negotiate missing item
Pay over the phone
Item wasn’t ready. Call back
Pick up

Book specialist (takes my insurance)

Research in-network providers
Call to verify insurance
Check availability
Submit new patient forms
Confirm appointment
Follow up on referral

Register kid for summer camp

Research camps & availability
Call about waitlist
Submit registration
Pay deposit
Confirmation never came. Follow up
Drop off forms

Dispute insurance bill

Find claims process
Call billing, wait on hold
Request itemized statement
File formal dispute
Follow up on status
Escalate if denied

Find lost luggage

Find airline claims process
Call baggage services
File claim
Submit receipt for reimbursement
Escalate after 48h
Pick up with claim ID

Plan event

Research venues & availability
Call to check capacity
Negotiate group rate
Submit deposit
Confirm catering & AV
Day-of coordination calls

Questions

What is Duckbill?

Duckbill is an execution API for AI agents. It gets real-world tasks done: phone calls, payments, errands, and services. Your agent sends the goal, Duckbill completes it and returns the outcome. The infrastructure behind it has handled 30M+ real-world actions.

Who is Duckbill for?

Two kinds of customers. AI agents and assistants that need reliable execution for their end users, and companies that want real-world tasks handled end to end without building an agent of their own. The first group integrates the capability APIs. The second uses Concierge.

How do I integrate Duckbill?

Through a REST API with webhooks, or through the Duckbill MCP server for Claude, ChatGPT, and agent frameworks. Most teams start with phone calls and add payments and errands as their users ask for them.

Does Duckbill use humans?

Every task gets the cheapest model or person that can complete it reliably, and fallback execution stands behind automation wherever it runs out. That is our problem, not yours. You ping the API. The task gets done.

How much does Duckbill cost?

A simple per-task fee. No specialists to hire, no BPO to manage, no ops team on your balance sheet. Tell us your volume and we will send pricing.

Don't become a services company.

Stripe meant you didn't have to become a payments company to take payments. Duckbill means you don't have to become a services company to get things done in the real world. No specialists to hire. No BPO to manage. Ping the API, pay per task.

Start with phone calls

REST API and MCP. Bring your own agent, or use ours.