rn devtools hub

Use it from an agent

Both plugins install the MCP server plus a skill that teaches the agent to chain its tools: check the project context before debugging anything that looks impossible, prove results with assertions instead of screenshots, wait on events instead of sleeping, and read an element's source instead of grepping the repository.

Claude Code

/plugin marketplace add rn-devtools-hub/rn-devtools-hub
/plugin install rn-devtools-hub

Codex

codex plugin marketplace add rn-devtools-hub/rn-devtools-hub
codex plugin add rn-devtools-hub

Then start the hub at the root of your app, which is what the agent talks to:

npx rn-devtools-hub

Registering the server by hand works too. Claude Code:

claude mcp add rn-devtools --transport http http://127.0.0.1:8973/mcp

Codex, in ~/.codex/config.toml:

[mcp_servers.rn-devtools]
url = "http://127.0.0.1:8973/mcp"

Cursor has no marketplace, so it is two files in your own project. Declare the server in .cursor/mcp.json:

{
  "mcpServers": {
    "rn-devtools": { "type": "http", "url": "http://127.0.0.1:8973/mcp" }
  }
}

And copy the rule, which is the Cursor equivalent of the skill:

mkdir -p .cursor/rules
cp node_modules/rn-devtools-hub/templates/cursor-rule.mdc \
   .cursor/rules/rn-devtools-hub.mdc

It ships with alwaysApply: false, so Cursor pulls it in when the task matches instead of paying for it on every request.

For any other agent, the skill on its own comes from the skills.sh registry:

npx skills add rn-devtools-hub/rn-devtools-hub

It writes .agents/skills/rn-devtools-hub and links it into whichever agents it detects, from Gemini CLI and Copilot to Windsurf and Zed. This installs the skill only, so register the server as above and start the hub to give the agent something to talk to.

Any client that speaks only stdio uses npx rn-devtools-hub mcp, which bridges to the hub and starts it on demand.

Screenshots

Why

The SDK lives inside the JavaScript runtime of your app.

An accessibility-driven tool sees what the OS exposes. A WebDriver-driven one sees a black box. An IDE inspector sees the tree but will not act on it. None of them can read a component's props, call a handler, intercept a request or write to a store, because none of them is inside.

That position is what the whole product is built on. Everything else follows from it:

The agent can	Because it is inside the runtime
Find an element by role and accessible name, then act on it	Actions go through the app's own props, not through pixels
Get the file and line that produced an element	The location lives in React's dev bookkeeping
Prove a step without a screenshot	A screenshot cannot show a request that failed silently
Tell a stale native build from a code problem	Only the runtime knows what the binary actually is
Freeze the clock and the network	Date and fetch are in the runtime
Put the app in an exact state without walking ten screens	Stores are reachable directly
Explain a visual regression, not just score it	The changed region maps back to the component that owns it

A pure JavaScript SDK in the app (inert in production), a local hub in a single process, a dashboard in the browser. Zero dependencies anywhere, including the hub. Your data never leaves your machine.

Features

Panel	What you see
Overview	KPI tiles, JS thread lag, HTTP statuses, duration distribution
Crashes	Fatal errors, JS errors, unhandled promise rejections, stacks
Network	Request/response inspector (colored methods, durations, sizes), copy as cURL, secrets redacted
Uploads	Live upload queue (if your app emits the events, see the protocol)
Cache	React Query snapshot: keys, statuses, freshness, data
Storage	AsyncStorage keys, sizes, values, live write timestamps
Screens	Navigation journey, time spent per screen
Logs	console.log/info/warn/error, colorized JSON, filters
SQLite	Read-only SQL console (SELECT/PRAGMA) on your app's database
Endpoints	Map of declared endpoints, calls, latencies
Actions	Buttons driving the app: reload, clear caches, your custom actions
Tests	Human-readable .hubflow scenarios, live replay progress, failure diagnosis, approved visual evidence and safe target repair candidates
Tools	What the agents do with this hub: calls per tool, failures with their message, empty answers and why, context bytes returned (pixels counted separately, because a session spending most of its context on screenshots is a session verifying the expensive way), selectors used, and the loop replayed against the app's own events
Plugins	The services around the app: App Store Connect and Google Play, what each one is configured with, whether it may change a release, and every host it will contact
Design	Icon, splash, fonts, sounds, identity (read from app.json and the assets)
Mirror	Live app screen (needs react-native-view-shot in the app), full Android via adb (tap, swipe, keyboard, Wi-Fi), iOS simulator via xcrun

The Overview panel opens with the project context: what the project declares, what the app actually runs, and whatever the two disagree on. A stale native build is the most common way to lose an afternoon here, and it is named before you start reading code.

Plus: multi-device with merged sessions, bug report export in Markdown (ready for a GitHub issue), real-time capability badges, and a local MCP server to drive everything from Claude, Cursor or any MCP client.

See Hubflow scenarios to record an exploration, save it as a causal E2E test, replay it in CI and inspect its selected success or failure screenshots in the dashboard.

What an agent gets

Tool	What it answers
get_project_context	What the project declares, what it actually runs, and the contradictions. Call it first when anything behaves impossibly
get_ui_tree, query_ui	The visible components, each carrying the source file and line that produced it
ui_act	Tap, type, submit, scroll, by element and never by pixel
assert	Proves a step: element kinds retry, event kinds catch a request that failed silently, a console error, an unhandled rejection
freeze_time, mock_network	Deterministic at the JS level, so a scenario runs the same twice
get_state, set_state	Put the app in an exact state without walking ten screens
render_component	Mount a component inside the running app, under its real providers
snapshot_baseline, compare_snapshot	A visual diff that names the component owning the changed region
export_session, export_flow	One correlated timeline, and actions paired with the consequences they caused
save_flow, list_flows, get_flow, run_flow, propose_flow_repair	Persist a causal scenario, replay it with assertions and screenshots, and create a reviewable repair candidate when a strongly identified target moves
audit_accessibility	What React renders but the accessibility tree does not expose
list_plugins, asc_*, gplay_*	The release around the app: has the build finished processing, where is the version in review, what is production serving, what do the store reviews say. And driving it: distribute to TestFlight, submit for review, promote a track, widen or halt a staged rollout
capture_store_screenshots	Regenerate the App Store and Play screenshots from the running app: devices, locales and screens from a manifest, each reached with the app's own dev actions, captured at native resolution and uploaded. The pixels never enter the agent's context

Source locations survive React 19, where the location lives in owner stacks pointing into the bundle: the hub symbolicates them against Metro before the agent sees them.

For AI agents, the hub also exposes runtime UI automation over MCP: get_ui_tree (semantic tree of the visible components, read from the React runtime, including native views without accessibility like maps), query_ui (find elements by role and accessible name, testID, placeholder, text or label, scoped with within, with measured rects), ui_act (tap, type exact text, submit, scroll, all by element, never by pixels, and it reads the field back to say whether the text landed), typed dev actions (list_actions/run_action: navigate, seed, login without touching the UI), and a correlated event flow (wait_for_event, get_events_since) that replaces sleeps with real signals like screen.ready. Works in Expo Go, development builds and bare React Native, and in CI without any simulator: enable it with devtools.attachUiAutomation().

The hub covers the native layer too, as a superset of what idb/simctl scripting gives an agent: session_start boots a dev build on the right Metro server with zero dialogs (permissions pre-granted, dev-menu onboarding skipped), plus set_permission, launch_app, open_url, screenshot_native, set_location, send_push, set_appearance, set_animations, set_overlay (the expo-dev-menu bubble covers native controls no UI tree can show) and the last-resort tap_native and swipe_native, on iOS simulators and Android devices, with every capability probed and degrading cleanly.

Everything above stops at the machine. The half of a release that does not is behind a plugin: App Store Connect and Google Play ship with the hub and stay inert until you give them credentials. They read a release and they drive one, so the agent that decides to widen a rollout is the agent that already has the crash evidence for it. Installing the hub's plugin installs the release skill with it, so there is nothing else to add.

A configured plugin prints the hosts it will contact at startup, an unconfigured one exposes no tool at all rather than costing an agent context it cannot use, and RN_DEVTOOLS_PLUGIN_WRITES=off removes every tool that could change anything. See docs/plugins.md.

Quick start

Prerequisites: Node 20+. Bun is used when present, and is not required. Optional capabilities (adb mirror, iOS simulator, Wi-Fi) have their own prerequisites: see the integration guide.

# 1. In your React Native / Expo project, as a devDependency:
npm install --save-dev rn-devtools-hub

# 2. Wire it up automatically (detects your libraries, writes the glue,
#    hooks the entry point, adds the `devtools` script):
npx rn-devtools-hub init

# 3. Start the hub (runs on Bun or Node 20+)
npm run devtools
# -> Dashboard: http://localhost:8973/?token=... (URL printed at startup)

init inspects your project and generates only the code it can actually run: axios interception if you use axios, the Storage panel if you have AsyncStorage, device info if you have expo-device, and so on. It never overwrites an existing glue file (use --force), and --dry-run shows what it would change.

That's it: logs, crashes and performance already flow in. Every additional integration (network, cache, storage, SQLite, mirror...) is a recipe of a few lines: see the integration guide.

Integration guide

The SDK is agnostic: it exposes generic primitives that you wire to YOUR libraries. All the recipes are in docs/integration.md, notably:

• devtools.attachAxios(instance, "api"): any axios instance
• devtools.wrapFetch(fetch, "uploads"): any fetch-based client
• devtools.emit(type, payload): feed any panel
• devtools.onCommand(name, handler): respond to the dashboard (e.g. SQLite)
• devtools.registerAction({name, label, danger, requiresNative}, handler)
• devtools.attachUiAutomation(): UI perception and actions for AI agents
• devtools.markScreenReady("Login"): "screen ready" signal agents wait on
• The complete events and commands protocol: docs/protocol.md. It is the contract: any tool (or any LLM) can integrate a panel by implementing it.

Host project dependencies

The SDK imposes NOTHING. Depending on the features you want, add to YOUR project:

Feature	Install in your project	Type
The package itself	rn-devtools-hub	devDependencies
App mirror (screen stream)	react-native-view-shot	dependencies (included in Expo Go)
Storage panel	@react-native-async-storage/async-storage	already present in most apps
SQLite console	expo-sqlite (or your driver + a sqlite.query handler)	depends on your app
Enriched device info	expo-device, expo-application, expo-network	dependencies
Full Android mirror	adb on the dev machine (not in the app)	system tool
iOS simulator mirror	Xcode command line tools (dev machine)	system tool

Security

• The SDK is inert outside __DEV__ (double guard: yours and the SDK's)
• The dashboard requires a token (printed at startup, RN_DEVTOOLS_TOKEN to pin it)
• Sensitive headers (Authorization, cookies, x-api-key) are redacted before leaving the device
• The SQLite console only accepts SELECT and PRAGMA
• The MCP server only listens on localhost, with Origin verification
• Design panel assets are confined to the project root, with whitelisted extensions
• Plugins are the only outbound path, they are inert without credentials, each one declares the hosts it contacts, and no credential is ever returned by a tool or written to a log
• Every plugin tool that changes something is declared as such, is annotated so MCP clients confirm it, and can be removed entirely with RN_DEVTOOLS_PLUGIN_WRITES=off
• RN_DEVTOOLS_SCREENSHOTS=off (or a number) removes or budgets pixel captures, for teams who do not want an agent verifying with screenshots

Contributing

Contributions are welcome. The full guide is in CONTRIBUTING.md. In short:

git clone https://github.com/rn-devtools-hub/rn-devtools-hub
cd rn-devtools-hub
npm install          # also installs the husky hooks
npm test             # vitest
npm run typecheck
npm run hub          # starts the local hub to test the dashboard

• Commits follow Conventional Commits (feat:, fix:, docs:...): the changelog and versions derive from them automatically (release-it), and commitlint checks them at commit time
• The pre-commit hooks run typecheck + tests
• Release: npm run release (maintainers)

For AI agents: read AGENTS.md and llms.txt.

License

MIT