Open source

React Native Agents

Twenty-four expert AI agents that give your coding assistant real React Native expertise instead of generic React advice — build failures, performance, security, code quality, accessibility, testing, native modules and release.

Works with Claude Code, Cursor, Windsurf, GitHub Copilot, Codex, Zed, Aider, any MCP client, and as a GitHub Action on your pull requests.

View on GitHub npm packageGitHub Marketplace

24

expert agents

128

reference docs

0

dependencies

MIT

licensed

The problem

Your AI assistant knows React. It does not know React Native.

So it reaches for useMemo on a primitive, suggests a web pattern that has no native equivalent, recommends Flipper five versions after it stopped mattering, or hand-edits ios/ in a managed Expo project where that directory is regenerated on every prebuild. React Native Agents replaces that guesswork with specialist playbooks that read your package.json first, cite real file:line locations, and tell you how to verify the fix.

The agents

Twenty-four specialists, 128 reference documents

Ask a normal question and the right specialist is selected automatically, or invoke one directly with its command.

Doctor

Day to day

Gradle, CocoaPods, Metro, Xcode, and environment failures — diagnosed from your actual error output

/rn-doctor

Build

Day to day

New screens, components, forms, and lists that already handle safe areas, accessibility, states, and keyboard

/rn-new

Debug

Day to day

Builds fine but behaves wrong — render loops, state that will not update, release-only bugs, post-Flipper tooling

/rn-debug

Dependencies

Day to day

Should we add this library — New Arch support, maintenance, native cost, lighter alternatives

/rn-deps

Onboard

Day to day

Orienting in an inherited codebase — architecture map, conventions, landmines, what to read first

/rn-onboard

Store Submission

Day to day

Rejection triage, privacy manifests, Data Safety, ATT, target API deadlines

/rn-submit

Monorepo

Day to day

Metro resolution across packages, hoisting, duplicate React, autolinking from a nested app

/rn-monorepo

Performance

Review

Lists, rendering, startup, memory, bundles

/rn-perf

Security

Review

Secrets, storage, auth, TLS, WebViews, deep links, supply chain

/rn-security

Code Quality

Review

Architecture, TypeScript, hooks, state, errors, RN idioms

/rn-review

UI & Accessibility

Review

Layout, safe areas, screen readers, contrast, RTL, theming

/rn-ui

Testing

Review

Jest, RNTL, mocking, Maestro, Detox, CI

/rn-test

Native Modules

Review

TurboModules, Fabric, JSI, codegen, Swift/Kotlin, packaging, bridge migration

/rn-native

Observability

Review

Crash reporting that silently doesn't work, symbolication, breadcrumbs, tracing, PII in telemetry

/rn-observe

Release

Review

EAS, signing, OTA, stores, rollout, rollback

/rn-release

Upgrade

Review

Version matrix, New Architecture migration, scope moves, and the regressions that compile cleanly

/rn-upgrade

Navigation

Review

Deep links, cold-start routing, auth guards, nested navigators, typed params

/rn-nav

Push

Review

APNs and FCM setup, token lifecycle, background handlers, tap routing

/rn-push

Permissions

Review

Denial and blocked states, purpose strings, rationale flows, partial grants

/rn-permissions

Platform Parity

Review

iOS/Android divergence — keyboard, safe areas, hardware back, shadows, pickers

/rn-parity

Offline

Review

Durable mutation queues, idempotency, optimistic rollback, conflicts, sync

/rn-offline

State

Review

Server vs client state, selectors and re-renders, persistence and hydration

/rn-state

Payments

Review

IAP and subscriptions, server-side receipt validation, restore, refunds, grace periods

/rn-pay

Background

Review

Background fetch, headless JS, location, and designing for the task not running

/rn-background

A further command, /rn-audit, runs every review specialist and consolidates the findings. Knowledge is verified through React Native 0.87 and Expo SDK 57 — always check version-specific advice against your own project.

Bundle size

Measured from the source map. No model call, no API key, no cost.

Bytes are attributed per source-map segment, so a package's number is measured rather than estimated. Because it is fully deterministic, it can run on every pull request for free. Budgets accept 250kb, 1.5mb or a raw byte count — an invalid value is an error, never a silently skipped check.

$ npx @maheshwarimrinal/react-native-agents size --base main --budget-delta 100kb

📦 Bundle size 📈 +332.0 KB
2.00 MB → 2.33 MB (+16.2%)

❌ Budget exceeded — increase 332.0 KB exceeds budget 100.0 KB

| moment 🆕 | +207.0 KB | 207.0 KB |
→ date-fns or dayjs (Hermes ships Intl, so formatting may need no library at all)

Design decisions

Built to be trusted on a real codebase

Knowledge that expires on purpose

A knowledge.json file records exactly what has been verified and when — React Native 0.87, Expo SDK 57, React 19.2. A scheduled workflow watches the npm registry and opens an issue when reality moves ahead of the docs. Stale advice becomes a bug report instead of a silent wrong answer.

Evals that punish cargo cult

Every agent ships with graded fixtures — realistic code plus what a competent engineer should say about it. Cases carry forbid rules too: the performance agent fails its own test suite if it reaches for FlashList before it has asked anyone to measure.

A failed run fails the check

Invalid key, no credits, provider outage — the audit exits non-zero instead of reporting a clean pass with zero findings. A green check that means "we did not actually look" is worse than a red one.

Routing that respects your bill

Seven agents are excluded from pull-request routing entirely — they need something brought to them: an error log, a question, a rejection notice. The rest are matched to changed paths, and signals are scoped narrowly enough that a typical UI change reaches about three agents rather than twenty-four. Adding specialists does not raise the cost of an ordinary pull request.

Telemetry off by default

Nothing is sent unless you turn it on, and even then it is package version, Node major, OS and which editor you installed for — never paths, repository names, code, findings or IP address. DO_NOT_TRACK is honoured and overrides an opt-in. Every field that can ever be transmitted is documented, and a test fails the build if that document falls out of sync with the code.

Zero dependencies

The frontmatter parser, the CLI and the MCP server are all implemented directly. Nothing in node_modules to audit, nothing in the supply chain to trust, and dist/ is committed so npx works without a build step.

Installs without clobbering

The installer previews changes before writing, skips conflicts by default, supports an explicit --force, and backs up user-authored files whenever it replaces one. Your existing rules survive the install.

Pull requests

Review every diff with the specialists that matter to it

The GitHub Action routes changed files to the relevant agents and posts findings to the pull request. Anthropic and OpenAI are both supported, with a per-run spend cap and configurable severity gates.

- name: React Native audit
  uses: maheshwarimrinal/react-native-agents@v1
  with:
    provider: anthropic
    model: claude-sonnet-5
    api-key: ${{ secrets.ANTHROPIC_API_KEY }}
    budget-usd: '1.50'
    fail-on: never

Store API keys as repository secrets — never commit them to a workflow file.

Supported tools

One source, generated for every assistant

Playbooks are authored once and compiled into each tool's native rule format, so the advice does not drift between the editor you use and the CI that reviews you.

Claude Code

Plugin, marketplace entry, subagents, slash commands

Cursor

.cursor/rules/*.mdc

Windsurf

Trigger-scoped .windsurf/rules/*.md

GitHub Copilot

Instructions, path-scoped rules, chat modes

Codex · Zed · Aider

AGENTS.md and .agents/react-native/

MCP clients

Zero-dependency stdio MCP server

Install it in one command

Free and MIT licensed. Issues, ideas and pull requests are all welcome.

Star on GitHubRead the docsSee the demo audit