Verification for coding agents
Ship agent code with evidence, not a green checkmark.
Omea boots your repo in a microVM, an agent authors and runs real Playwright tests against it, and you only keep the ones that prove coverage moved.
$ npx @omea/cli run
✓ booted your app in an isolated VM
✓ ran your suite with coverage instrumented
! checkout > applies discount code — failed
evidence: .omea/last-run/coverage.mdHow it works
Agents write code. Omea verifies the work.
Connect GitHub. Your agents verify their own work — and run 24/7. Omea boots your app in an isolated VM and hands your coding agent a verdict it can act on mid-task, not a green checkmark it has to take on faith.
A read-only GitHub App connects your repo — you pick which ones, we never push. That connection is coming soon; today, point the CLI at a local checkout instead.
Your app boots in an isolated VM with dependencies, environment, and mocks wired automatically, talking to platform twins like a fake Shopify instead of production keys. Boot-and-run is shipped locally today; our agent writing the initial tests and connecting twins in the cloud is next.
Run it locally right now and get a pass/fail verdict with evidence, not a claim to trust. The cloud loop — verdicts against your current workspace at high concurrency, so agents can keep going around the clock — is what's coming next.
Shipped today: npx @omea/cli run runs entirely locally — no login, nothing leaves your machine. Coming soon: npx @omea/cli login and the hosted loop that runs the same verdicts against your current workspace from our infrastructure, 24/7.
The trust recipe
You don't have to trust the agent — only the evidence it leaves behind.
The only access we ask for is a read-only GitHub App scope on the repos you choose. We never push, and there's no write scope to revoke later.
Every run leaves a trace you can open yourself — coverage report, pass/fail per test, logs — not a self-reported "tests passed." Today that lands locally under .omea/last-run/.
Tests are plain @playwright/test files committed to your repo, never a proprietary format. Delete the connection and npx playwright test still runs the day after.
Why this exists
You already know the two ways this breaks.
"Tests passing" stopped being evidence. An agent can quietly weaken an assertion, comment out the part that fails, or fake the terminal output and say the work is done — a green checkmark from the same agent that wrote the change isn't proof of anything.
Worktrees isolate code, not the environment. Run a few coding agents in parallel and they're still fighting over the same port, the same database, the same state — Omea forks a VM per test, so parallel runs never collide.
Engineering evidence
Numbers we can point to, not numbers we assert.
913ms cold vs 64ms warm, mean latency to a verification session's first response, isolating the harness's own overhead from LLM latency. Measured in omea-fleet/docs/pre-warm-pool-latency.md.
The CLI's workspace sync and the cloud's run protocol are built independently and compose byte-for-byte end to end. Proven in omea-cli/scripts/cloud-roundtrip.sh.
No customer counts, no savings claims on this page — we don't have them yet, and we won't invent them.
Built for the agent calling it
What npx @omea/cli run --json hands back to your agent.
Every run streams NDJSON your agent can act on without human interpretation: test-started/test-finished events as tests complete, then one terminal run-finished event carrying a versioned report — exit status, per-test results, a coverage summary with the delta against the last run (did the test your agent just wrote actually move coverage), and the paths to every artifact. Shipped today in the local CLI as schema_version 2. The hosted loop (coming soon) reuses the identical shape, adding only a shard_id — nothing an already-integrated agent needs to change.
Indicative pilot
Start with one page of scope, one verified loop.
$2.5–4K/mo
Pilot band for early teams proving the verification workflow. This will change as packaging hardens, so we are not over-building pricing yet.
No pSEO, docs, or blog in this cutCompare vs QA Wolf · Compare vs testRigor · Compare vs Ranger · Compare vs Momentic · Compare vs Bug0 · Compare vs Spur