AI customers for developers: an email API, a phone number and a browser for your agent

An email API for AI agents, such as AgentMail, gives each agent its own inbox, and an AI customer needs that plus a phone number, a browser, a sign up program for each platform, waiting that costs nothing for months, and an evidence store that keeps each email with its DKIM result and a hash. You can build it from services such as AgentMail, Twilio, Browserbase and Cloudflare Durable Objects, or call a hosted system over MCP and an API.

Engineers who build agents to track rivals, research prospects or test their own sign up flow need the agent to be a customer: to receive the welcome email, the login code by text and the price change 3 weeks later, and to show what arrived in a form anyone can check.

Each part exists as an API. AgentMail gives agents their own inboxes to send, receive and act on email, Twilio leases phone numbers that receive texts, and Browserbase runs browsers that agents drive. The work is in joining them into a customer that lasts months at thousands of companies, at almost no cost while it waits, with evidence that holds up.

What you are building

An AI customer is an AI agent that becomes a real customer of a company, with its own inbox, phone number and browser, and keeps what the company sends and shows it as evidence. In an API, each one is returned as a customer agent: 1 company, 1 identity, and the evidence it has kept. The definition and its neighbours are in AI customers.

A customer lasts months where an agent session lasts minutes, so most of the engineering moves from compute to state: an identity that keeps working, waiting that costs nothing, and evidence that still checks out after the sender has replaced its keys. A system that does this has 5 parts.

1. Identities: an inbox per company, a phone number per market and a browser.
2. Sign up programs: 1 per platform, learned once and replayed as code.
3. Durable waiting: state that sleeps until an email, a text or a timer wakes it.
4. An evidence store: each email and page as it arrived, with its time, DKIM result and hash.
5. Typed facts: 1 sentence each, with a link to its evidence, sent where they are used.

The build, part by part

Identities: an inbox, a phone number and a browser

Give each AI customer its own inbox at each company it joins, so every email belongs to 1 company and 1 watch. An email API for AI agents creates one in a single call. AgentMail’s pricing lists 10 inboxes for $20 a month on Developer, 150 for $200 on Startup, and $2 a month for each extra inbox. Gmail’s tabs are Gmail’s own sorting, so if you need to know whether an email landed in Promotions, that inbox has to be a Gmail mailbox.

A phone number per market receives texts and login codes. Twilio lists a US long code at $1.15 a month and $0.0083 per SMS segment sent or received, plus carrier fees, and says US A2P 10DLC numbers are subject to registration onboarding fees. Take each text by webhook and file it under the company that sent it.

A browser opens pages, fills in forms and reaches checkout. Browserbase describes itself as a platform for agents that browse and interact with the web like humans, and its pricing includes 100 browser hours on the $20 Developer plan, then $0.12 a browser hour.

Every identity is made for the job and belongs to whoever runs it. It never uses a real person’s name, address or card.

Sign up programs, 1 per platform

An AI customer signs up through the forms any customer uses: the pop up, the footer form, the checkout tick box, the trial form. With double opt in, a subscriber clicks a link in a confirmation email or text before being added, so a sign up is finished only when the inbox has received that email and the link has been opened.

Most companies collect sign ups on a few email and form platforms, so treat each platform as 1 program. Work out the first sign up on a platform with an agent, record the steps, and replay them as code at every company built the same way. Keep the agent for forms no program recognises, and version every program, because a page change can break one without raising an error.

Identify your agent, and stop where a site says no. Cloudflare’s Web Bot Auth lets bots and agents sign their requests with HTTP message signatures, and Shopify’s Storefront API gives signed traffic higher rate limits than unsigned bots. An AI customer never gets round a CAPTCHA or a bot wall. The report names the step it couldn’t test.

Durable waiting

Much of what a customer receives arrives late: the reminder the next day, the win back email after 60 days, the renewal notice after 11 months. A process kept running for each customer would bill for every hour of that wait, so keep each AI customer as state that sleeps and wakes on 3 events: an email, a text or a timer.

On Cloudflare, Durable Objects that are idle and eligible for hibernation are not billed for duration. Requests, which include alarm invocations, come with 1 million a month on the Workers Paid plan and then cost $0.15 a million, and that plan has a minimum charge of $5 a month. An AI customer that wakes 30 times a month makes 30 requests.

The evidence store

Store each email exactly as it arrived, in an append only store: the raw message, its headers, the time it landed and a hash of its bytes. Keep each page capture the same way, with its address and time. Verify the DKIM signature on arrival, and keep the public key you verified it with.

Arrival is the moment to check. RFC 6376, the DKIM standard, lists message archiving among the things it is not designed for, and describes how a domain replaces a key by publishing a new selector and then removing the old public key. A genuine email checked a year later can fail only because its key is gone. A result recorded on arrival, kept with the key and the hash, still shows which domain signed the message and that it hasn’t changed since.

Typed facts

Turn each piece of evidence into facts with a type, a company, a time, 1 sentence and a link back: welcome_email.arrived, code.refused, price.changed. Use plain code where the format is fixed, such as a price on a page, and a small model where it isn’t, such as telling a welcome from a reminder. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens, so classifying 30 emails of about 4,000 tokens each, with a short answer for each, costs about 2 cents.

Version the rules that turn evidence into facts. When a rule changes, derive the facts again from the stored evidence instead of visiting the company again, because last month’s emails can’t be collected twice.

The hard parts at scale

For 5 companies, a developer can join these parts in a few days. At thousands, most of the work sits in 7 places. The engineering of each, and the durable execution engines that handle the waiting, is in long running AI agents, and 10 systems to build on them are in what to build with AI customers.

• The long tail of forms. A few platforms cover most sign ups. The rest each have their own form, their own confirmation step and their own changes after every release.
• Pages that change. A replayed program can break when a field moves, so each one needs a check that it still works and a fall back to an agent.
• Waiting for months, with timers that survive deploys and wake ups that run once, even when an event arrives twice.
• Evidence that outlives key changes, checked on arrival and stored with its key and hash.
• Telling messages apart. A welcome, a reminder, a price change and an order confirmation look alike to a parser, and the categories differ by sector.
• Countries. Prices, offers and delivery messages differ by market, so each market needs its own number and address.
• History. What a company sent last month can’t be collected today. Sharing 1 AI customer per company across every user who watches it gives a new user that history on day one.

The rules hold at any scale. An AI customer says it is an AI whenever it talks to a person, uses no fake payment details, never gets round a bot wall and never clicks a rival’s paid ads at volume, because Google Ads treats automated clicks as invalid.

Build or buy

Obsession makes AI customers, so every other option is described from its provider’s own pages as of 24 October 2026.

Each service does its part well. You still write the sign up programs, the waiting, the evidence store and the facts. To put a number on the parts, count your companies and multiply: 1 inbox each, 1 number per market, about 10 browser minutes, and 30 wake ups and 30 classified emails a month.

In this example the inboxes are 89% of the bill, and the services come to about $1.68 per company a month. The larger cost is the engineering time to write and keep the sign up programs, the store and the classifier, so set your own hours against a hosted system’s price. The full cost model, against people and mystery shoppers, is in AI customers at scale.

Build it when you watch a handful of companies, need the store in your own systems, or want to run an open source runtime yourself. Buy when you need thousands of companies, history from the first day, or numbers in many countries.

Calling AI customers over MCP and an API

MCP is an open source standard for connecting AI applications to external systems. Ecommerce MCP servers such as Shopify’s act for a buyer: through them, agents search products, build carts, check out and fetch an order’s status. A customer-side server reports what each company sent and showed its customers, with the evidence.

Obsession exposes AI customers as 5 verbs over MCP, an HTTP API and TypeScript and Python SDKs: lookup (which companies the shared record already covers, and since when), facts (what a company did, each with its evidence), evidence (the raw email, page or transcript behind a fact), become_customer (start customer agents, or join ones already there) and notify (new facts pushed to your webhook, Slack or CRM as they land). Its runtime is open source: it fans out to many companies at once, sleeps each customer agent for free and learns each sign up once before running it as code, and Obsession Cloud runs it with inboxes and numbers in many countries and months of shared history. Keys open from a waitlist, priced per call or per company watched with a spending cap you set, and an agent connects to the MCP server at mcp.useobsession.com.

Limits

• A bot wall or a CAPTCHA stops it, and the report names the step it couldn’t test.
• A new customer sees what new customers get. Long standing customers may receive other emails and other inbox placement.
• It records outputs. Why a company sent something sits in that company’s own systems.
• Some journeys need a real purchase, on a card its owner funds.
• You may not need it. For a few companies checked now and then, a spare inbox, a phone and a weekly script are enough.

Questions from developers