npx @kifas/mcp@latest

Agents write your code. Kifas tests it.

One MCP server, 65 tools, first-class in Claude Code, Cursor, and Windsurf. Your agent drives real browsers, iOS and Android emulators, and a Device Farm of physical devices, and the loop covers your repo on every push.

$10 in credits, no card. Cold starts land in under three seconds.

~/acme-web · claude-code
// Install Kifas's MCP server
$ npx @kifas/mcp@latest --install cursor
→ Installed 65 tools. Authenticated as maya@acme.dev
 
claude > "Test our checkout with a new Gmail user, apply promo GUMBO20, verify the total."
 
→ Kifas: workflow.create("checkout-promo")
→ Kifas: workflow.add_node(type: "auth", primitive: "gmail")
→ Kifas: workflow.add_node(type: "action", ai_explore: true)
→ Kifas: workflow.run() · 14 nodes · 3.1s start · passed
 
// Compiled to /tests/checkout-promo.spec.ts
One tool surface

Real browsers, emulators, and physical devices.

Chromium · Gmail + TOTPiPhone 16 · Stripe checkoutPixel 9 · Team inviteChromium · Agent eval

Cloud browsers · Emulators · Device Farm · p50 < 3s cold start

The gap AI opened

The agent wrote the feature in twenty minutes. Nobody tested it. Code velocity tripled the day you adopted Cursor. Your verification did not get the memo.

What your agent gets

Give the agent that wrote it a way to prove it.

Kifas is a testing platform your agent talks to natively, plus a loop that keeps covering the repo when no one is at the keyboard.

01
MCP-native, end to end

65 tools, one server, npx @kifas/mcp@latest. Built around Model Context Protocol from the first call, not bolted on after.

02
Everything the agent can drive

Cloud browsers (Chromium, Firefox), iOS and Android emulators, and a Device Farm of real physical devices. All callable from the same tool surface.

03
The loop runs on every push

Scan, suggest, build, run, self-heal. Coverage keeps pace with commits even when no human is watching the pipeline.

04
Named, persistent state

Save a logged-in session or a seeded workspace as a named handle and reuse it across multi-turn agent runs instead of rebuilding it every time.

And it runs on its own, on every push:

01
Scan
read the repo
02
Suggest
tests worth having
03
Build
write and run them
04
Run
on every push
05
Self-heal
heal or flag
The 65 tools

Not a number. A surface.

65 tools, grouped into the families your agent calls. A few from each below; the full reference, with arguments and examples, lives in the docs.

Workflow
  • workflow.create
  • workflow.run
  • workflow.diff
  • workflow.publish
+ more in the reference
Auth + state
  • auth.add
  • auth.rotate
  • state.save
  • state.load
+ more in the reference
Devices
  • browser.spawn
  • device.spawn
  • device.tap
  • device.install
+ more in the reference
Run control
  • run.read
  • run.replay
  • run.flake_score
  • run.compare
+ more in the reference
Inboxes
  • inbox.create
  • inbox.wait_for
  • inbox.list
  • inbox.purge
Tunnel
  • tunnel.start
  • tunnel.attach
  • tunnel.list
  • tunnel.stop
Explore the full tool reference →
Fair questions

Native, and yours to drive.

Is this a wrapper, or actually MCP-native?

Built MCP-first. The dashboard you see is a UI over the same 65 tools your agent calls. There is no second, lesser API hiding behind the pretty one.

Loop, or standalone?

Either. Every piece runs standalone from your agent, and the autonomous loop is the fast path on top when you want coverage to run itself. Your call on who holds the wheel.

What does an agent run cost?

A credit is about a penny. Cloud browsers run at 2 credits a minute, mobile emulators at 5, and the Device Farm of real devices at 22. AI steps are 1 to 12 credits depending on depth. The free tier's $10 covers real usage, and adding teammates costs nothing.

Priced by what you actually run.

Credits, not seats. Adding teammates costs nothing. You spend when tests run.

Free
$0
$10 in credits · no card · 14 days
  • Cloud browsers, emulators, Device Farm
  • MCP server and 65 tools
  • Community support
Start free
Pro
$199/mo
The autonomous loop, self-healing, and the full device range.
  • GitHub-connected autonomous loop
  • Self-healing with break thresholds
  • Auth Catalog · vibe testing · hybrid engine
Choose Pro

See all plans →

Hand the keyboard to your agent.

npx @kifas/mcp@latest

Start freeRead the MCP docs