Polar as merchant of record: what metered AI billing actually commits you to
Polar — A billing platform for the intelligence era
At a glance
- What is it?
- Polar is a hosted billing platform that sells itself as the merchant of record, meters tokens, agent steps and GPU seconds, and charges a percentage plus flat cents per transaction. The repository is precise about rates, SDKs and workloads, and silent on exports, fee floors and the retry window.
- Who is it for?
- Polar fits teams that want usage billing without writing receipts, dunning and VAT logic themselves. It fits badly if your own customer database has to stay canonical, or if you need a contractual answer on downtime, refunds and data handling.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- polarsource/polar10,487 ★ · +213
| Repository | 2026-09-22 | 2026-11-15 |
|---|---|---|
| polarsource/polar | 10.3k | 10.5k |
213 Stars in the last 54 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Polar issues the receipt under its own name
Polar's central claim is commercial rather than technical: it will be the merchant of record. That means Polar takes on receipts, customer accounts, dunning, sales tax and VAT, so a product can charge for tokens, agents and compute without building billing infrastructure from scratch. The practical shape is that the charge a customer sees is issued by Polar, and the customer account that owns it is created on Polar's side rather than in your database. Tax is presented as worldwide and handled, with checkout deciding where the customer is instead of you shipping a tax engine. What that arrangement does not hand you is any documented way out. The README does not document exporting customer records, moving a subscription to a different account, or what happens to stored customer data when the relationship ends. The consequence for readers is concrete: if your product keeps its own users, you now hold two records of the same person, so settle the export and deletion path before you push the first live payment through it.
Four pricing tiers, and the flat cents that ride on every charge
The rate card is a percentage plus a per-transaction amount, and moving up the ladder only shrinks the second term slowly. Starter is free at 5.00% plus 50 cents per transaction, Pro is 3.80% plus 40 cents, Growth is 3.60% plus 35 cents, and Scale is 3.40% plus 30 cents. The same page also says additional fees may apply and links to a separate fees document, so those four headline rates are not the whole bill. Two things are simply absent: no minimum charge is stated, and nothing explains how the per-transaction amount interacts with the Startup Program, which gives eligible startups a year on the most generous plan. Model the flat cents against your average charge size before picking a tier. For sub-dollar events it becomes the dominant cost, not the percentage.
Metering reaches tokens, agent steps and GPU seconds
The billing units are unusually concrete for a billing product: tokens, API calls, agent runs, GPU seconds and storage, metered with precision down to the event. AI completions are described as wrapped with metered token billing on every call. Autonomous agents are charged per run with step-level granularity, and retries are handled without double billing. Fine tuning, hosted inference and training runs are metered by the second. That retry promise is the one a reader should slow down on, because it moves where your idempotency work lives. What the documentation does not say is how a retry is recognized as the same attempt, how long that recognition lasts, or what a provider-side retry after a timeout looks like when it never reached Polar at all. There is also no documented way to void or correct a meter event after the fact. For anyone charging per agent step, that means reconciling against your own provider logs, not just handling a webhook.
Two official SDKs, one API surface, and 1.0.1 framework packages
Integration runs through a Public API and a separate Webhook API, with endpoints documented for the latter. First-party SDKs exist for two languages only: JavaScript for Node.js and browsers in polarsource/polar-js, and Python in polarsource/polar-python. The three most recent releases listed are @polar-sh/[email protected] and @polar-sh/[email protected] (both 2026-10-30) and @polar-sh/[email protected] (2026-10-16). The consequence for readers is direct. Any runtime outside those two SDKs, which in practice means Go, Rust, Java or PHP back ends, means generating a client from the API or writing raw HTTP calls yourself. The credit list shows the project already depends on openapi-typescript-codegen to produce its own clients, so that route has precedent.
Running the platform yourself is a workspace, not a package
If you expected to install Polar, the repository will disappoint you, and it is honest about that. There is no install command and no environment variable list; contributors are pointed at DEVELOPMENT.md for the development environment. What the root does show is a workspace spread across two runtimes. server/, clients/ and sdk/ hold the application, infra/, terraform/ and lambda/ hold deployment, docs/ and handbook/ hold documentation, and dev/ sits alongside them. Tooling versions are split the same way: a Brewfile for macOS packages, a .nvmrc for Node, flake.nix and flake.lock for Nix, pyrightconfig.json for Python type checking, and a polar.code-workspace for editors. The root package.json carries no dependencies at all and only pins the package manager:
{
"private": true,
"packageManager": "[email protected]"
}Getting the source is the only step the repository asks for, and it hands you an environment specification rather than a deployable service:
git clone https://github.com/polarsource/polar.gitThat is the honest cost line for self-hosting: two runtimes and an infrastructure layer. Nothing here documents a production deployment path for operators.
Checkout geolocates customers through an outside IP data provider
Worldwide tax is sold as handled, and the mechanism behind customer location at checkout is credited to IPinfo, which supplies the IP address data. For a checkout flow that is a real external dependency, and it is worth naming it as one rather than as a detail. What the documentation does not describe is the failure path. It says nothing about what checkout does when the lookup fails, when the address it returns is wrong, or when the customer sits in one country while the billing address is in another. There is no statement about letting a customer correct the tax region by hand. The consequence for readers is that the fallback lands on your own support process. Test a checkout where IP geolocation and billing country disagree before you send real traffic, because no correction route is described.
Bugs go through a dashboard button, disclosures through SECURITY.md
Support has two documented doors and neither is a public tracker. Feature requests and bugs go through the Feedback button in the sidebar, bottom left of the Polar dashboard.. Security problems go to SECURITY.md, which the README names as the route for responsible and private disclosures. CODEOWNERS sits in the repository root, so a set of owners is defined, though ownership is not a response time. The repository is not archived and its last push was on 2026-11-02, so nothing here suggests abandonment. The gap is commitment rather than activity: no uptime promise, no support tiers, no refund handling statement and no data processing terms appear anywhere in the repository description. Ask for those in writing before you put a paid product on top of it.
Editorial conclusion
Polar fits teams that want usage billing without writing receipts, dunning and VAT logic themselves. It fits badly if your own customer database has to stay canonical, or if you need a contractual answer on downtime, refunds and data handling. Before you route live payments through it, check three things: the export and deletion path for customer records, the per-transaction cents measured against your average charge size, and whether an SDK exists for the language your backend actually uses. If your metering depends on retries or provider-side redelivery, reconcile Polar's meter events against your own logs before you trust the invoice.
Frequently asked questions
What kind of company is Polar in this repository?
It is the product behind the polarsource GitHub organization: a billing platform described as a billing platform for the intelligence era, licensed Apache-2.0, written mainly in Python, sold as a merchant of record and hosted at polar.sh.
What does the name Polar mean here?
In this repository the name refers to that billing platform: usage billing, subscriptions, seats, credits, trials and discounts, with metered charges for tokens, API calls, agent runs, GPU seconds and storage. The README does not say why the product was named Polar.
How much does Polar keep on each transaction?
Every tier is a percentage plus a flat amount: Starter 5.00% plus 50 cents, Pro 3.80% plus 40 cents, Growth 3.60% plus 35 cents, Scale 3.40% plus 30 cents. Additional fees may also apply, and eligible startups can apply to the Startup Program for a year on the most generous plan.
Can I run Polar myself instead of using the hosted service?
Nothing in the README offers an install command. Contributors are sent to DEVELOPMENT.md for the development environment, and the repository root carries server/, infra/, terraform/, lambda/, a Brewfile, a .nvmrc and flake.nix, which reads as a two-runtime workspace rather than a package you add to a project.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/polarsource-polar)Star history chart for your README
[](https://hysenlabs.com/projects/polarsource-polar)Was this useful?