What's new in Kifas.

The GitHub-connected autonomous loop, plus self-healing tests.

Live now. GA, not beta. This is the big one. Connect a GitHub repository and Kifas now runs the entire testing loop for you.

The loop

• Scan. Agents read your codebase and map the routes, components, and user flows worth covering.
• Suggest. You get a ranked list of end-to-end tests, critical paths first.
• Build. Agents author the tests and run them on Kifas cloud browsers, mobile emulators, and Device Farm devices.
• Run. Tests run on every push, with p50 cold starts under three seconds.

Self-heal

Your test code lives on the Kifas platform, so when a test fails, Kifas can compare it against the code change and work out what kind of failure it is. A moved selector or a renamed label gets healed on the spot, and you get a notification showing exactly what changed. A genuine behavior change is treated as a real bug: it breaks the connected GitHub Action job, drops a comment on the PR, and can fire a webhook to kick off anything else you have wired up. You decide how easily a test is allowed to break, or leave it on our default.

Composability is unchanged

The MCP server, the 42 tools, manual control over every browser and device, and standard Playwright export are all still here. The loop is the fast path. You can still drive any piece by hand.

To turn it on: connect GitHub from your dashboard, pick a repo, and let the first scan run.

Device Farm: real devices, on demand.

Mobile testing now comes in two tiers. Emulators stay at 5 credits/min and cover the bulk of your flows. The new Device Farm gives you real physical devices at 22 credits/min for the things only real hardware can prove, like camera, biometrics, GPU, and hardware-specific bugs. It is the same SDK call with one extra argument.

Auth Catalog adds more providers.

More one-line login flows, including additional SSO providers. Pick the flow, pass test credentials, skip the OAuth scripting.

Faster cold starts.

Browser spawn time is down to roughly 820ms, with p50 cold start under three seconds. Three flake fixes shipped alongside, all from customer reproductions.

Hybrid engine, officially stable.

The headline change: ai.step now compiles to deterministic Playwright on the first successful run, then replays at Playwright cost forever after. We've been dogfooding this since September; it's been driving most of our internal suites without a regression for six weeks.

Added

• Hybrid engine GA. ai.step auto-locks after first pass.
• Self-heal: locked steps fall back to AI on failure, re-lock, post a Slack diff.
• Per-step budget cap: ai.step(goal, { maxBudgetUSD }).
• iOS 18.2 added to the Simulator fleet.
• 3 new Auth Catalog providers: WorkOS, Stytch, Clerk.

Changed

• Default Chromium target moved to chromium-131.
• workflow.run() now streams step events; old polling endpoint is deprecated (still works for 90 days).
• MCP server --install now auto-detects your active client (Cursor / Claude Code / Windsurf).

Fixed

• Race in state.load when state and target browser were in different regions.
• Visual regression false-positives when an iframe rendered after the snapshot finished.

MCP server gets six new tools.

Most of these come from agent feedback. Claude kept asking for capabilities we'd assumed lived elsewhere. We listened.

Added

• Total MCP tools: 42. (We're aiming for 50 by GA.)
• Per-tool latency now visible in the MCP card via run.read.

iOS Simulators · beta opens.

Six months of macOS-fleet plumbing later: you can now spawn an iPhone 16 simulator with one SDK call. Pricing is unchanged; iOS sims share the same browser-minute pool.

Added

• device.spawn('iphone-16', { os: '18.2' }): beta on Team and above.
• WebInspector bridge for hybrid native + WebView tests.
• Region: us-east-1 only at beta; eu-west-1 in v1.4.

Changed

• Median spawn time across the fleet down to 820ms (was 1.4s).

Pricing simplified.

We retired the per-seat upcharge that snuck back in over the summer. Unlimited seats are baked into every paid tier again, full stop. If you noticed your Sept invoice climb, you'll see it drop in November.

Changed

• Per-seat fees on Starter and Team removed retroactively for Q4.
• Pricing page now shows the in-page calculator we should have shipped at launch.

Workflow canvas, version one.

The visual workflow editor is out of preview. Drag-to-connect, three-way diff with code-authored edits, and round-trip JSON serialization to your repo.

Added

• Visual canvas at app.kifas.io/wf/:id.
• Sub-workflows pinned by SHA in CI.
• Branch / loop / retry / parallel control nodes.

Stability week.

Three flake fixes that came out of customer reproductions.

Fixed

• Cookie partitioning bug on WebKit 17.6.
• auth.loginAs('okta') failed when push-notif was the only available second factor.
• Tunnel reconnect storm when a CI runner cycled between jobs in <3s.

Auth Catalog · 42 providers.

The Auth Catalog grew from 12 to 42 providers, with first-class flows for TOTP, push, SMS, magic link, and recovery codes. Each one runs in an isolated browser per project, with cookies encrypted under a project-scoped KMS key.

Kifas, generally available.

After 14 months of beta with 200+ teams, Kifas is GA. Three primitives, one SKU, MCP-native from day one.

Added

• Workflows, state handles, devices: the three primitives.
• SDKs for TypeScript, Python, Go.
• Playwright fixtures via @kifas/playwright.
• SOC 2 Type II report available on Business+.