Choose what
needs to work.
Start with a customer task, your site, and a clear definition of success.
GOALComplete checkout as a guest
Agent compatibility testing
Can agents
complete your
checkout?
Test real tasks across the AI agents your customers use. See every failure on record. Ready for the agentic web.
Free to start. Usage included.
Clarity, from the inside out.
THE QUESTION THAT MATTERS
Can an agent complete the task?
Your website. A new kind of visitor.
Find a product.Book a stay.Complete checkout.
Put your site to the test
The task is
the test.
Give agents a real goal. See whether they can reach it. When something breaks, go straight to the evidence.
Compatibility checkTASK / COMMERCE
Find the blue tote and complete guest checkout.
3 example profilesSame goal. Different journeys.
RUN REPLAYExample profile B
COMMON GOODSCheckout
04 / 04
From a question to a clear answer
A new test.
A familiar workflow.
Choose what
Run it across
your agent matrix.
Put the same task in front of different agents. Repeat runs to see how consistently it works.
RUNS✓✓✓AGENTS✓—✓
Find the failure.
Check the fix.
Inspect the replay, give your team the evidence, and test again after a change.
NEXTReplay → Improve → Rerun
The agent matrix
Test the agents
that matter.
Every profile names the agent, the model version, and how it is accessed. API-direct runs and emulated profiles are always labeled — evidence you can stand behind.
Claude computer use
Anthropic's vision agent, driven step-by-step with full reasoning capture.
Gemini computer use
Google's browser-control model, coordinates grounded against your live pages.
OpenAI computer use
The Operator-class model via the Responses API.
browser-use + GPT
The most-deployed open agent framework — the developer long tail hitting your site today.
Consumer products
Real-product runs (agentic browsers and assistants) for ground-truth calibration, clearly distinguished from API profiles.
Scripted baseline
A deterministic control run, so agent variance is measured against a known-good path.
The release gate
Your PR broke checkout
for an agent. Know first.
Run the matrix against every preview deploy. Results land beside your existing tests — JUnit output, commit statuses, webhooks. An MCP server, so your own agents can drive Flightbox too.
$ flightbox ci --suite checkout \
--base-url $PREVIEW_URL --junit results.xml
[PASS] checkout--browser-use-gpt--1 6 steps $0.02
[FAIL] checkout--claude-cu--1 (grounding) replay ↗
exit 1 — blocking merge
Confidence that keeps up
Your site changes.
Agents change.
Keep it working.
Build a baseline for your critical journeys. Compare results as your site evolves, and bring agent compatibility into your release process.
Plan your test coverageMetered. No gates. No tiers to outgrow.
Pay for runs.
Nothing else.
Every feature from the first dollar — all agents, CI gate, MCP, replays, reports. You pay per recorded run, like tokens. Spend caps are yours to set.
FREEStart here
Your first
clear answers.
$10 free credit on signup
Enough for a full multi-agent matrix on your checkout, judged and recorded.
- All agents, all features, day one
- Replays, transcripts, friction telemetry
- No card to explore
METEREDPay as you go
Published rates.
Automatic discounts.
from $0.05 per recorded run
One run = one task × one agent × one attempt, fully recorded.
- Scripted $0.05 · DOM agents $0.40 · vision agents $1.50
- Bring your own model keys: from $0.10 platform fee
- 10–30% off automatically at volume
MONITORED JOURNEYOptional
Set it. It watches.
$200/journey/mo · pilot pricing
A scheduled matrix run on one critical journey — baseline, trends, and an alert when an agent version changes. Includes its own run credits; overage just meters.
Failed on our infrastructure: not billed. Failed by the agent: billed — the evidence is the product. Prepaid credit packs and a $7,500 founder-led compatibility audit available.
Find your starting point
A little testing.
A wider view.
See how your tasks, agents, and test frequency add up.
ESTIMATED VOLUME72test attempts / month
3 tasks × 3 agents × 2 attempts × 4 weeks
A planning estimate, not a bill. A month is modeled as 4 weekly or 30 daily checks. Volume discounts apply automatically from 1,000 runs/month.
A few useful answers
Before your
first run.
What does Flightbox test?
Whether an AI agent can complete a task on your website under the conditions you define. That could be finding a product, choosing options, completing checkout, booking a stay, or creating an account.
How is this different from a browser test?
A scripted browser test follows steps you specify. An agent test starts with a goal and lets the agent decide how to reach it. Flightbox is designed to show whether that journey succeeds across different agents and repeated attempts.
Does one successful run mean my site is ready?
A successful run is evidence for that task, agent, and environment. Repeat attempts and compare agents to understand consistency. Keep testing as your site and agent versions change.
How does free usage work?
You get $10 of run credit on signup — enough for a judged multi-agent matrix on one journey. After that it is pay-as-you-go at published per-run rates, with per-run, per-suite, and monthly budget caps you control. Runs that fail on our infrastructure are never billed.
Which agents can I test?
Today: Claude computer use, Gemini computer use, OpenAI computer use, and browser-use with GPT-class models — all API-direct, plus a scripted control. Every profile is labeled with the agent, model version, and access method; consumer-product calibration runs are on the roadmap and will be labeled distinctly. The console above is illustrative — the recorded demo batch is real.
The next way in is already taking shape.
Be ready
to let it in.
Your website. Ready for agents.
