rn-dev-agent

v0.67.3 · Claude Code + Codex · iOS + Android

Your agent writes the code.
This proves it runs.

rn-dev-agent turns Claude Code and Codex into a React Native development partner. It reads the component tree, store state, and navigation stack over CDP, taps real UI on iOS and Android, and replays verified flows in seconds.

/plugin marketplace add Lykhoyda/rn-dev-agent
Get started →
  1. ❯ /rn-dev-agent:test-feature "cart badge updates on add"
  2. ⚙ cdp_status ✓ connected · Metro :8081
  3. ⚙ cdp_navigate Products ✓ 96 ms
  4. ⚙ device_press @addToCart ✓ 142 ms
  5. ⚙ cdp_component_tree ✓ CartBadge count=3
  6. ⚙ cdp_store_state cart ✓ items: 3 · total: $47.00
  7. ⚙ maestro_run cart-badge ✓ replayed in 4.2 s
  8. ✓ Verified on iPhone 16 Pro · action saved: add-to-cart
210×action replay speedup
79MCP tools
3–25 mindescription → verified feature
2platforms — iOS + Android

The problem

Coding agents ship blind.

They write the feature, satisfy the type-checker, and declare victory — without ever seeing the app run. "It should work now" is not verification.

The fix

Close the loop on device.

After implementing, the agent connects to your running app, navigates to the screen, reads the component tree and store state, exercises the real UI, and only then reports done — with the evidence attached.

Three layers, one contract

Introspect. Interact. Replay.

L1 Introspect

Read the running app's truth over Chrome DevTools Protocol — component tree, Redux/Zustand/React Query state, navigation, network, logs.

cdp_store_state cart
→ { items: 3, total: 47.00 }

L2 Interact

Tap, type, swipe, and scroll through in-tree native runners — XCTest on iOS, UIAutomator2 on Android. Self-building, self-healing.

device_press @checkoutBtn
→ ✓ 142 ms · settled

L3 Replay

Verified walks are saved as parameterized actions and replayed as prologues — a 14-minute interactive flow reruns in about 4 seconds.

cdp_run_action login
→ ✓ replayed in 3.8 s

See the output

Tools that return evidence, not vibes.

{ "testID": "cartBadge",
  "type": "Badge",
  "props": { "count": 3, "visible": true },
  "children": [{ "type": "Text", "text": "3" }] }
{ "store": "cart", "kind": "zustand",
  "state": { "items": 3, "total": 47.0,
             "lastAdded": "sku_2481" } }
{ "ok": true, "ref": "@addToCart",
  "meta": { "timings_ms": { "resolve": 18, "tap": 96, "settle": 28 },
           "keyboardGuard": "no_keyboard" } }
{ "action": "add-to-cart", "status": "passed",
  "durationMs": 4180, "autoRepair": "not-needed",
  "runRecord": ".rn-agent/actions/runs/2026-07-10.json" }

/rn-feature-dev

From description to verified feature.

  1. Discover
  2. Explore
  3. Question
  4. Architect
  5. Implement
  6. Verify live
  7. Review
  8. E2E proof

Phase 6 is the difference: the agent drives the simulator, confirms the UI and state, and saves the verified walk as a replayable action for next time.

Free · open source · MIT

Stop taking your agent's word for it.

/plugin marketplace add Lykhoyda/rn-dev-agent
Read the docs →