SparkTech · SparkTest

Your tests write themselves.
When they break, they heal.

Describe a test in plain English. SparkTest turns it into a real, auditable Puppeteer script — then a fleet of AI agents on our own silicon runs it around the clock and repairs it when your UI changes.

Start free trial See how it works

14 days · 5 maintained tests · no card required

sparktest console · run #4128

PASSED · healed

describecodifyexecutehealreport

▸claim run RUN-4128 · storefront / checkout

▸script v3 cached · guest-can-reach-payment

·launch chromium · headless · 1440×900

·goto https://shop.acmeweb.dev

✓add "Aurora Lamp" to cart

✓cart badge shows 1

✕click #checkout-btn — selector not found

⟳heal · error + DOM → local LLM · patching v3 → v4

✓click [data-test=checkout] — recovered

✓payment page reached

▸report · 9 steps · 41s · 1 heal · script v4

agent mini-04 · local LLM · outbound HTTPS

healing diff ready for review →

93%

Less test maintenance

Self-healing removes the selector-chasing that eats QA time.

~€0

Marginal run cost

Owned, energy-efficient silicon. Electricity, not metered tokens.

24/7

Autonomous fleet

SparkNodes polling for work around the clock.

100%

Auditable artifacts

Real Puppeteer scripts — readable, versioned, exportable.

How it works

Five steps. Zero test maintenance.

SparkTest is not a recorder and not a prompt wrapper. It is a pipeline that turns intent into living, auditable automation.

01 · Describe

Write test cases in plain English, organized by Project → Stream → Test. “A guest can add a product to the cart and reach payment.” That is the spec.

02 · Codify

On first run, an agent turns the description into a real Puppeteer script — a readable, versioned artifact stored with the test case. No black boxes.

03 · Execute

The fleet runs your scripts headless against the target site, capturing logs, timings and screenshots. On schedule, on commit, or on demand.

04 · Heal

When a run fails, the agent feeds the error and DOM context to a local LLM, patches the script and re-runs — bounded attempts, full diff, humans in the loop.

05 · Report

Pass rates, healing events, duration trends and flakiness — rolled up per stream and project in the SparkTest Console.

Broken tests stop being tickets. They become diffs you approve.

Architecture

Built like infrastructure, not like a demo.

A cloud control plane and an execution fleet, separated on purpose. Your engineers can audit every layer — and every script.

Outbound-only agents

Agents enroll with a one-time token, then poll for work over outbound HTTPS. No inbound ports, no VPN, nothing to open in your firewall.

Versioned Puppeteer artifacts

Every codify and every heal produces a new script version with a reviewable diff. Export them anytime — they run anywhere Node runs.

Metadata-driven platform

The control plane runs on the SparkTech Backbone platform: entities, workflows and permissions expressed as governed configuration, not one-off code.

The fleet

Our moat is boring: we own the hardware.

Cloud-LLM testing tools leak money on every run, so they meter you. SparkTest runs on purpose-built AI nodes with local models — which changes the economics, and therefore the product.

Owned silicon

The fleet is SparkNodes — compact AI appliances we own and rack ourselves, not cloud instances we rent. Inference costs electricity, not tokens — so a test run is essentially free, forever.

Purpose-built nodes · CapEx not OpEx

AI Powered

Local LLM on every agent

Codify and heal happen on-device. Your DOM, your flows and your failures never leave the fleet — no third-party model API in the loop.

On-device inference · data stays in fleet

Deterministic capacity

No metered throttling and no per-run anxiety. Reserved agent slices on Scale, named machines on Dedicated Fleet — hosted or on-prem.

Pause · drain · retire from the console

Runs cost us electricity. So we don’t meter yours. Run on every commit — that’s the point.

Self-healing

Your UI changed overnight. Your tests already know.

Industry surveys put 30–50% of QA engineering time into fixing broken selectors — not finding bugs. SparkTest gives that time back.

The agent feeds the error and DOM context back to a local LLM, patches the script and re-runs — bounded attempts, never an infinite loop.
A healed test is reported as healed, with the exact script diff. You review the change; you never write it.
Healing is included in every plan. You never pay to fix a broken test — that is the product.

checkout/guest-can-reach-payment v3 → v4

healed · run #4128

@@ step 3 · proceed to checkout @@

await page.goto(BASE + "/cart");

await page.waitForSelector(".cart-list");

-await page.click("#checkout-btn");

+await page.click("[data-test=checkout]");

-await page.waitForSelector(".spinner", { hidden: true });

+await page.waitForSelector("[data-test=payment-root]");

await assertHeading(page, "Payment");

2 attempts · patched in 12s · re-run green

✓ passed (healed)

Pricing

Don’t pay for seats or runs. Pay for living tests.

The unit of value is a maintained test: a test we keep alive, healed and reporting. Healing is included in every plan.

Starter

€149/mo

A living smoke suite for one product.

25 maintained tests
1 project
Daily schedule + on-demand runs
Fair use 1,500 runs/mo
Shared agent pool
Self-healing included
Email support

Start free trial

Stop maintaining tests. Start shipping.

14-day trial · 5 maintained tests · shared fleet · no card required.

Start free trial See pricing

Your tests write themselves.When they break, they heal.