Privacy Policy

Who we are

Margin (“Margin”, “we”, “us”) builds a cross-provider tool that measures what AI-agent work costs per unit of output. This policy covers trymargin.io and the Margin product and APIs that link to it.

Margin is based in the United States and the product runs on US infrastructure. You can reach us at any time at subh@trymargin.io.

Two different kinds of data, governed differently

Most privacy policies describe one relationship. Margin has two, and conflating them would hide the part that matters most to a technical buyer.

• Information we control: what we collect when you visit this website, email us, or ask to be told when Margin opens up. We decide why and how this is used, and this policy governs it.
• Customer data we process: the call and outcome records a customer’s systems send to our ingest API. The customer decides what is in it; we act on their instructions. That relationship is governed by our Data Processing Addendum, not by this policy. The section below on what the meter collects is here so you can see the shape of it before you sign anything.

Information we collect

Information you give us

• Email correspondence: the site’s only conversion action is a mailto link. If you email us, we hold your address, your name if you give it, and whatever you write, for as long as the conversation and any resulting relationship is live.
• Waitlist: our API accepts a waitlist signup consisting of an email address, an optional name, and which page it came from. No tracking identifier, no IP address, no device fingerprint is stored with it.

Information collected automatically on this website

We use PostHog (US cloud) for product analytics and session replay. On each page view it records:

• the URL, path and page title, the referring URL and domain, and the search engine you arrived from, if any;
• browser name and version, operating system and version, device type, browser language, screen and viewport dimensions;
• an approximate location derived from your IP address:country, region, city, postal code, timezone, and an approximate latitude and longitude with an accuracy radius. This is coarse, and it is derived at the point of collection;
• a randomly generated identifier for your browser, held in a first-party cookie and in local storage, so repeat visits can be counted as one visitor;
• clicks and other interactions on page elements, including clicks that hit nothing and repeated clicks in one spot, which is how we find broken or confusing parts of the page;
• a session replay: a reconstruction of your visit showing mouse movement, scrolling and clicks.

We also use Sentry for error monitoring. When the site throws an error, Sentry receives the error and stack trace along with the URL and browser details. It is configured not to attach personal information, and errors are sampled rather than sent in full.

What the meter collects from a customer’s systems

This is the part worth reading closely, because it is the opposite of what people expect from a tool that sits near their AI calls.

Margin is a meter, not a proxy. Your agents call your model providers directly. Our SDK measures the call after the fact and posts a record of its economics. It is not in the request path, it cannot read a request, and our ingest going down cannot stop your agents.

One exception, and you start it. When you connect a GitHub repository in the console and press Execute, we clone it and run the eval harness it contains in a throwaway container on our own server, paying for the model calls through a key made for that run alone, capped in dollars and revoked when the run ends. Those calls, prompts and responses included, go from our server to OpenRouter. The container and the checkout are deleted when the run ends; a daily server backup taken during a run can hold a copy for up to 7 days. We keep the run’s call and outcome records like any other.

A call record consists of:

• identifiers and labels you choose: workflow_id, session_id, prompt_id (a label for a prompt, not the prompt), operation, environment, team, harness, task_key;
• what was called: provider and model name, whether the call was a retry, whether it ran on a batch tier, why the generation stopped;
• what it cost: input, output, cache-read, cache-write and reasoning token counts, computed cost, and latency in milliseconds;
• structure: span identifiers linking a call to the agent that spawned it, the orchestration pattern, branch width, stopping rule;
• outcomes: whether a unit of work passed, a quality score and how it was graded, and optionally a link you supply pointing at the artifact.

How we use information

• To run and improve the site and the product: which pages get read, where people give up, what breaks.
• To answer you when you email us, and to have the conversation that follows.
• To provide the product to a customer: computing cost-per-outcome, comparing routes, and producing the readouts the customer asked for, on that customer’s instructions.
• To keep things secure and working: diagnosing errors, investigating abuse, enforcing rate limits.
• To meet legal obligations and to establish or defend legal claims where we have to.

We do not train models on customer data. Customer call and outcome records are used to produce that customer’s own results. We do not use them to train or fine-tune any model, and we do not pool one customer’s data into a benchmark shown to another. The DPA states this as a binding commitment.

We do not sell or share personal information, including as “sale”, “sharing” and “targeted advertising” are defined under California, Colorado, Connecticut, Virginia and comparable state laws. We run no advertising and no cross-context behavioural profiling. We have never done either.

Who we share it with

We use a small number of service providers to run Margin. They process data on our instructions and for no purpose of their own.

We also use Anthropic, Google and OpenRouter for Margin’s own inference: grading outcomes and running our own analysis. To be exact about what that means: these are providers we call in the course of operating Margin. They are not a route your production traffic takes, and your production prompts never reach them through us, because we never have them. The one exception is a measurement run you start from the console, described above: its model calls go to OpenRouter from our server, and DigitalOcean hosts the server they run on.

Beyond that, we may disclose information if the law requires it, to protect our rights or someone’s safety, or to a party that acquires the business, in which case this policy continues to apply to information transferred until it is replaced by one you are told about.

How long we keep it

• Session replays: 30 days, then deleted automatically by PostHog.
• Website analytics events: no longer than 24 months.
• Email and waitlist entries: until you ask us to delete them, or until the conversation has plainly ended and there is no reason to keep it.
• Customer call and outcome records: for as long as the customer’s agreement is live. On request, or within 30 days of the agreement ending, we delete them or return them, at the customer’s choice.

We would rather hold less than more. If you want something deleted sooner, ask and we will do it.

Security

Traffic to the site and the API is encrypted in transit. Ingest keys are stored as SHA-256 hashes, never in plain text, so a copy of our database does not yield a working key. Each project’s records are tagged server-side with the authenticated project, so one customer cannot write or read another customer’s data.

What we do not claim: Margin holds no SOC 2 report, no ISO 27001 certificate and no HIPAA attestation. We are a small company and we would rather say that than display a badge we have not earned. No system is perfectly secure, and we do not promise one.

Your choices

• Do Not Track: turn it on and we collect no analytics from your visit. Most privacy policies tell you the opposite; we honour it.
• Cookies: see the Cookie Notice for what is set and how to clear or block it.
• Email: we send no marketing email. If we ever do, every message will carry an unsubscribe link.
• Deletion: email us and ask. You do not need to invoke a statute for us to honour it.

US state privacy rights

Depending on where you live (California, Colorado, Connecticut, Virginia, Utah, Texas, Oregon, Montana and a growing list of others), you may have the right to know what personal information we hold about you and where it came from, to get a copy of it, to correct it, to delete it, and to opt out of its sale, its sharing for cross-context behavioural advertising, or profiling with legal effects.

We do not sell personal information, do not share it for cross-context behavioural advertising, and do not profile anyone in a way that produces legal effects. So there is nothing to opt out of. The other rights are real and we will honour them.

To exercise any of them, email subh@trymargin.io. We will ask for enough information to be confident you are who you say you are. Usually replying from the address in question is enough. We will not treat you differently for asking. If we turn a request down, you may appeal by replying to our answer, and we will explain the outcome in writing.

California residents may also ask, once a year, for a list of personal information disclosed to third parties for their direct marketing purposes. We disclose none, so the answer is nil.

If a customer’s records are involved, the customer is the controller and we will refer your request to them. See the DPA.

Notice to people in Europe and the UK

Margin is based in the United States with no establishment in the EEA, Switzerland or the UK. Where the GDPR or UK GDPR applies, we are the controller of the information described in this policy, and a processor for customer records under the DPA.

Our legal bases:

• Legitimate interests: running and improving a website, keeping it secure, and answering people who contact us. We have weighed this against your interests, which is part of why the site stores no IP addresses and masks everything typed into a replay.
• Consent: for non-essential analytics and session replay where consent is required. You can withdraw it at any time by clearing site data or enabling Do Not Track; withdrawal does not undo processing already carried out.
• Contract: to provide the product to a customer.
• Legal obligation: where we have to.

You have the right to access, correct, delete, restrict and object to processing of your personal data, to receive it in a portable form, and to complain to your supervisory authority. Ask us at subh@trymargin.io.

International transfers: our infrastructure is in the United States, so personal data collected in Europe or the UK is transferred there. Where we rely on a transfer mechanism we use the European Commission’s Standard Contractual Clauses, with the UK Addendum and the Swiss adaptations where those apply. The clauses populated for customer data are attached to the DPA.

Other sites, and children

This site links to places we do not run: PyPI, npm, GitHub, the sources behind figures we cite. Their privacy practices are their own.

Margin is a tool for engineering teams. It is not directed at children, and we do not knowingly collect personal information from anyone under 16. If you think a child has given us information, tell us and we will delete it.

Changes

If we change this policy we will update the date at the top. If a change matters, such as a new category of data, a new purpose or a new processor, we will say so here rather than quietly reissuing the page.

Contact