agent-react-devtools
CLI that connects to a running React or React Native app via the React DevTools protocol and exposes the component tree, props, state, hooks, and profiling data in a token-efficient format.
Core Workflow
- Ensure connection — check
agent-react-devtools status. If the daemon is not running, start it with agent-react-devtools start. Use agent-react-devtools wait --connected to block until a React app connects.
- Inspect — get the component tree, search for components, inspect props/state/hooks.
- Profile — start profiling, trigger the interaction (or ask the user to), stop profiling, analyze results.
- Act — use the data to fix the bug, optimize performance, or explain what's happening.
Essential Commands
Daemon
agent-react-devtools start # Start daemon (auto-starts on first command)
agent-react-devtools stop # Stop daemon
agent-react-devtools status # Check connection, component count, last event
agent-react-devtools wait --connected # Block until a React app connects
agent-react-devtools wait --component App # Block until a component appears
Component Inspection
agent-react-devtools get tree # Full component hierarchy (labels: @c1, @c2, ...)
agent-react-devtools get tree --depth 3 # Limit depth
agent-react-devtools get component @c5 # Props, state, hooks for a specific component
agent-react-devtools find Button # Search by display name (fuzzy)
agent-react-devtools find Button --exact # Exact match
agent-react-devtools count # Count by type: fn, cls, host, memo, ...
agent-react-devtools errors # List components with errors or warnings
Performance Profiling
agent-react-devtools profile start # Start recording
agent-react-devtools profile stop # Stop and collect data
agent-react-devtools profile slow # Slowest components by avg render time
agent-react-devtools profile slow --limit 10 # Top 10
agent-react-devtools profile rerenders # Most re-rendered components
agent-react-devtools profile report @c5 # Detailed report for one component
agent-react-devtools profile timeline --limit 10 # First 10 commits (use --limit; uncapped can dump 300+ lines)
agent-react-devtools profile timeline --limit 10 --offset 10 # Next 10 (pagination)
agent-react-devtools profile timeline --sort duration --limit 5 # Top 5 most expensive commits
agent-react-devtools profile timeline --sort timeline --limit 5 # Explicit chronological order (same as default)
agent-react-devtools profile commit 3 # Detail for commit #3
agent-react-devtools profile export profile.json # Export as React DevTools Profiler JSON
agent-react-devtools profile diff before.json after.json # Compare two exports
Understanding the Output
Component Labels
Every component gets a stable label like @c1, @c2. Use these to reference components in follow-up commands:
@c1 [fn] App
├─ @c2 [fn] Header
├─ @c3 [fn] TodoList
│ ├─ @c4 [fn] TodoItem key=1
│ └─ @c5 [fn] TodoItem key=2
└─ @c6 [host] div
Type abbreviations: fn = function, cls = class, host = DOM element, memo = React.memo, fRef = forwardRef, susp = Suspense, ctx = context.
Components with errors or warnings show annotations: ⚠2 = 2 warnings, ✗1 = 1 error. Use agent-react-devtools errors to list only affected components.
Inspected Component
@c3 [fn] TodoList
props:
items: [{"id":1,"text":"Buy milk"},{"id":2,"text":"Walk dog"}]
onDelete: ƒ
state:
filter: "all"
hooks:
useState: "all"
useMemo: [...]
useCallback: ƒ
ƒ = function value. Values over 60 chars are truncated.
Profiling Output
Slowest (by avg render time):
@c3 [fn] ExpensiveList avg:12.3ms max:18.1ms renders:47 causes:props-changed changed: props: items, filter
@c4 [fn] TodoItem avg:2.1ms max:5.0ms renders:94 causes:parent-rendered, props-changed changed: props: onToggle
Render causes: props-changed, state-changed, hooks-changed, parent-rendered, force-update, first-mount.
When specific changed keys are available, a changed: suffix shows exactly which props, state keys, or hooks triggered the render (e.g. changed: props: onClick, className state: count hooks: #0).
Common Patterns
Wait for the app to connect after a reload
agent-react-devtools wait --connected --timeout 10
agent-react-devtools get tree
Use this after triggering a page reload or HMR update to avoid querying empty state.
Diagnose slow interactions
agent-react-devtools profile start
## User interacts with the app (or use agent-browser to drive the UI)
agent-react-devtools profile stop
agent-react-devtools profile slow --limit 5
agent-react-devtools profile rerenders --limit 5
Then inspect the worst offenders with get component @cN and profile report @cN.
Browse a long timeline in chunks
agent-react-devtools profile timeline --limit 20 # commits 0–19
agent-react-devtools profile timeline --limit 20 --offset 20 # commits 20–39
agent-react-devtools profile timeline --offset 30 --limit 10 # skip warm-up, show 30–39
Use profile commit <N> to drill into a specific commit once you spot a spike.
Find a component and check its state
agent-react-devtools find SearchBar
agent-react-devtools get component @c12
Verify a fix worked
agent-react-devtools profile start
## Repeat the interaction
agent-react-devtools profile stop
agent-react-devtools profile slow --limit 5
## Compare render counts and durations to the previous run
Using with agent-browser
When using agent-browser to drive the app while profiling or debugging, you must use headed mode (--headed). Headless Chromium does not execute ES module scripts the same way as a real browser, which prevents the devtools connect script from running properly.
agent-browser --session devtools --headed open http://localhost:5173/
agent-react-devtools status # Should show 1 connected app
Important Rules
- Labels reset when the app reloads or components unmount/remount. After a reload, use
wait --connected then re-check with get tree or find.
status first — if status shows 0 connected apps, the React app is not connected. The user may need to run npx agent-react-devtools init in their project first.- Headed browser required — if using
agent-browser, always use --headed mode. Headless Chromium does not properly load the devtools connect script.
- Profile while interacting — profiling only captures renders that happen between
profile start and profile stop. Make sure the relevant interaction happens during that window.
- Use
--depth on large trees — a deep tree can produce a lot of output. Start with --depth 3 or --depth 4 and go deeper only on the subtree you care about.
References
| File |
When to read |
| commands.md |
Full command reference with all flags and edge cases |
| profiling-guide.md |
Step-by-step profiling workflows and interpreting results |
| setup.md |
How to connect different frameworks (Vite, Next.js, Expo, CRA) |