Write browser tests in plain English, keep the full power of TypeScript, and run the whole suite on every pull request for dollars a month instead of thousands.
// tests/checkout.test.ts
import { faker } from "@faker-js/faker";
import { test, expect, secret } from "sedum-cli";
test(
"a customer checks out",
{ url: "https://www.saucedemo.com/" },
async ({ page, ai, env }) => {
await ai.group(
"Log in",
[
"type {{user}} in the username field",
"type {{password}} in the password field",
"click the login button",
],
{ user: "standard_user", password: secret(env.SAUCE_PASSWORD!) },
);
await ai("click the Add to cart button for {{product}}", {
product: "Sauce Labs Backpack",
});
await expect(page.locator(".shopping_cart_badge")).toHaveText("1"); // plain Playwright
await ai("click the shopping cart link");
await ai("click the Checkout button");
await ai("type {{first}} in the First Name field", {
first: faker.person.firstName(),
});
await ai("type {{last}} in the Last Name field", {
last: faker.person.lastName(),
});
await ai("type {{zip}} in the Zip/Postal Code field", {
zip: faker.location.zipCode(),
});
await ai("click the Continue button");
await ai("verify the order summary lists Sauce Labs Backpack");
},
);npx sedum run tests/checkout.test.tsEach ai(...) sentence is one step that Sedum resolves on the page, with no
selectors. Everything between steps is ordinary code: Playwright's page and
expect, test data from any library, API calls to seed or skip setup, and
helper functions.
Sedum uses a model for two jobs only: finding the element a sentence refers to, and judging whether a claim such as "a list of products with prices is shown" holds on the page. That model is Jev, from TypeSafe. It returns probabilities, not free text. Everything else is deterministic: clicking, typing, waiting, verdicts, and exit codes run on Playwright.
Status: pre-alpha. The first alpha is on npm as
sedum-cli. The test format may change before 1.0, and Windows is experimental. Linux and macOS are verified.
You need Node.js 20.19 or newer and a TypeSafe API key or a compatible provider key.
mkdir my-sedum-tests && cd my-sedum-tests
npm init -y
git init # locator results are cached in Git metadata
npm install -D sedum-cli
npx sedum init # an example test, config, and .env.example
npx sedum browsers install chromium # if init says it is missing
cp .env.example .env # then set TYPESAFE_API_KEY in .env
npx sedum run tests/example.test.ts --headedFor a compatible provider, set TYPESAFE_BASE_URL to its API root and
TYPESAFE_API_KEY to that provider's key. Set TYPESAFE_DEFAULT_MODEL if its
model name differs from the default. See
provider configuration.
The example signs in to a demo shop. To see a failure, change its last step to
a false claim, such as verify an error message says the password is incorrect,
and run it again: it fails with exit code 1. If anything is missing,
npx sedum doctor says what and how to fix it.
New to AI browser tests? Read plain-English browser tests first.
Writing tests in English is not new. Running them on every pull request is usually too expensive, because hosted tools charge per step. Sedum is open source, runs on your machine with your own key, and calls a model that is fast and cheap enough to ask on every step.
The cost row is 50 tests of 10 steps (3 of them AI steps), run about 20 times a day by a 20-person team, priced at a per-step platform's published pay-as-you-go rates and at Jev's posted token price ($91 with the cache off). The speed row is five runs each with every cache off, on one laptop. See sedum.dev for the full method.
How it compares with other ways to test in a browser:
- Probabilities, not guesses. Every claim is scored against a threshold you set, and checked for contradicting evidence. A marginal pass is flagged, not silently green. See probabilistic testing.
- No black box. Prompts, scoring, and caching are in this repo. You can see exactly what goes to the model and what comes back.
When a test fails, have Sedum write a report for an agent such as Claude Code or Cursor:
npx sedum run --reporter markdown
# markdown .sedum/runs/<run-id>/report.mdThen ask the agent to read that file and fix the failure. The report puts the most urgent problem first, with the sentence and its file and line, the scores, the page text the model judged, the elements it considered, a screenshot, and a command to rerun just that test, so the agent can check its own fix. Page text is fenced and labelled untrusted, so a page cannot inject instructions into the report.
A test is a test() in a *.test.ts file. Its body gets ai for
plain-English steps, and Playwright's page and context for everything else.
Write {{name}} in a sentence and pass the value separately, so the sentence
stays the same on every run and can be checked before a run. Wrap passwords
and tokens in secret(): they are typed, but never sent to the model or shown
in reports. A file can hold several tests, and each gets a fresh browser
context. See TypeScript tests for the full API, and
examples/saucedemo for a small suite that uses faker,
skips the login form with a cookie, and checks app state behind the UI.
Tests that need no code can also be *.test.yaml files with a list of
sentences. Both formats run side by side in one project.
# tests/login.test.yaml
url: https://www.saucedemo.com/
data:
user: standard_user
password: $SAUCE_PASSWORD
steps:
- type {{user}} in the username field
- type {{password}} in the password field
- click the login button
- verify a list of products with prices is shown$VAR reads an environment variable and is treated as a secret. Put shared
steps in a *.module.yaml file and call it with use:. See the
YAML format.
- TypeScript tests: test(),ai, values and secrets, groups,extract, Playwright between steps, and results.
- YAML format: data, secrets, modules, beforeandaftersteps, ids, and tags.
- CLI commands: filters, parallel runs and sharding, retries, timeouts, reporters, and exit codes.
- Running in CI: GitHub Actions, GitLab, and Jenkins, with JUnit and job summaries.
- Project configuration: sedum.config.yaml, environments, thresholds, and.env.
- Probabilistic testing: how scores become verdicts, flags, and exit codes.
- Privacy: exactly what is sent to the model, and what never is.
- All documentation
Recently shipped: tests in TypeScript, the first alpha on npm, sedum init to
scaffold a project, parallel runs and sharding, and JUnit reports.
Planned for the 0.1 alpha:
- a GitHub Action
- starting your app before a run
State the outcome and let Sedum work out the actions, then independently verify it:
url: https://shop.example.com/
goal: >
sign in, add a hat to the cart, check out with the saved card
verify: the confirmation page shows an order numberGoal tests are YAML only for now. Use goal and a required verify claim instead of steps. Supply typing values
through data; before and after hooks remain available. Run and validate the
file with the same CLI commands as authored-step tests. See goal mode
for budgets, reporting, supported operations and recorded limitations.
Issues and pull requests are welcome. See CONTRIBUTING.md to build from source, and SECURITY.md to report a vulnerability. Sedum is MIT licensed.