devtools CLI

The devtools command line tool reads what the devtools recorded in a running app, and can mock, block or send a request in it. This page lists everything it takes and prints. Use with AI agents shows how to use it, with a snippet for your agent.

Running it

bun devtools runs a devtools.ts file first if the folder has one. In a monorepo, or outside the app's folder, npx devtools can miss the local install and fetch an unrelated npm package named devtools. npx @axonpack/expo-devtools <command> always runs this one. Below, devtools stands for whichever form you use.

Commands

apps and topics are probes: the AI Usage tab doesn't list them. In apps, the devtools column is the package version in the app, - when the app doesn't have it, and ? when it can't be checked without --force.

Flags

A flag is --name value or --name=value. A value can start with a dash, as in --sort -ms. -h is the only short flag.

Query flags

The same on every topic. get takes --full, --json and --out.

App flags

Needed only when more than one app is running, or for an app on another port.

apps takes --port, --force and --json. topics takes the app flags and --json. An action takes the app flags and its action flags.

Where clauses

<field><op><value>, or <field>?.

A plain value with commas is any of them: method=get,post. A /regex/ is never split on commas. An object value is matched as its JSON.

Times

--since and --until take a number and a unit, ms, s, m, h or d (30s, 1.5h), counted back from now on the device's clock, or any date the JavaScript Date reads, such as an ISO time. A row with no time, such as every storage row, is left out by either.

Topics

id and t are on every topic: id is what get, --after and --before take, and t is the time --since reads, printed as an ISO time. Fields in bold are the columns shown by default. A --fields name that doesn't exist answers with the topic's fields.

network

HTTP requests, WebSockets and server-sent event streams since launch, the last 1,000. get adds a socket's frames as messages and a stream's events as events. A request's initiator lists each frame as a file with its line and column, and keeps frames from libraries only with --full.

--where state=error lists every request that failed: a 4xx or 5xx answer, a request that got no answer, and a socket that failed. --where 'code>=400' lists only the 4xx and 5xx answers. A socket has no code; how it closed is in close.

console

Console lines since launch, the last 500. A line repeated in a row is one row with a count.

performance

Long tasks, slow interactions, marks and measures, newest first. Kept only while the Performance tab records, and it starts paused. Memory and frame-rate samples are not rows. An empty list while the tab is paused ends in # paused: the Performance tab records only after record is pressed.

storage

Every key of every registered store, by store then key. Rows have no time. Each query lists every store's keys again, so keys the app wrote since show up, and keeps the values already read. A list then reads the values on its page, and --grep or a --where, --sort or --group-by on value, type, size or error reads every key once. A value read earlier is not read again.

query

The TanStack Query cache: queries and mutations, the one updated last first. off with no QueryClient in config.query.client.

navigation

The moves the navigator made, the last 200. off with no Expo Router or React Navigation found.

redux

Dispatched actions, the last 200, each with what it changed. Never the whole state. off with no store using devtoolsReduxEnhancer().

get adds each change: its path, whether it was added, removed or changed, and the value before and after.

crash

Captured crashes, newest first. get adds the stack, the breadcrumbs and the device. off when crash.enabled is false.

ai-usage

The queries and actions agents made, the last 100, each with the reply it got. A query of ai-usage keeps no reply. --where 'action?' lists the actions alone.

Actions

devtools <topic> <action> [args] [flags] changes something in the app. Only network has actions; devtools topics lists each topic's actions. Query flags (--where, --json, --out and the rest) don't go with an action, and action flags don't go with a query: either is an error before the tool looks for an app. Every action, refused ones too, is listed in the AI Usage tab, marked as an action.

Action flags

network actions

Overrides match the exact URL the app requests, scheme and query included, as the Network tab's own overrides do. There are no patterns or prefixes: copy the URL from a row. They are the same overrides the tab makes, so they show in its Overrides view and in React Native DevTools. They apply to fetch and XMLHttpRequest, and last until the app reloads or they are removed.

send sends the request from inside the app, through the same function as the Network tab's Sandbox, and so through the app's own fetch. Overrides and network conditions apply to it, and it is a row in the log like any other request. The method is case-insensitive.

replay takes a logged HTTP request and sends it again as the Sandbox would: the same method, URL, headers and body, with --header and --body on top. The app's own headers go with it, including ones the tool shows you only as [redacted]. A header that network.redactHeaders removed before the request was logged goes as the text [redacted]; pass it again with --header.

Both answer with the new row, in the network columns plus error:

# network · my-app · iPhone 16 Pro · sent POST https://api.example.com/v1/notes as fetch-1762172310455-215. devtools get network fetch-1762172310455-215 shows it whole
id	method	code	state	ms	url	error
fetch-1762172310455-215	POST	201	success	312	https://api.example.com/v1/notes	-

replay starts replayed <id> as <new id>. A request that failed (blocked, offline, no answer) is still a row, with state error, and exit code 0. When the request is not in the log (recording is paused, or network.redact dropped it) the reply is # network · sent <METHOD> <url>: <status> in <ms> ms. Not in the log: recording is paused, or the app's network redact config dropped it, and its body is not shown.

While the request is in flight the app answers # wait: sending <METHOD> <url> as <id>, and the tool asks again every 200 ms within the 5 seconds. A request that takes longer carries on in the app. Running the command again would send it again, so the error points at ai-usage instead, where the action is listed once it ends.

Output

Tab-separated, with a header row. In a cell, - is an empty value, a tab or line break is written as \t, \n or \r, t is an ISO time, and a string over 2,048 characters ends in ... [cut at 2048 of <n> chars, --full for all].

--json prints none of the # lines, so it has no cursors. Each row is one JSON object with the chosen fields; --group-by gives { "<field>": value, "count": n }. Strings are still cut unless --full. With no rows it prints nothing.

Errors

The app's own errors are one error: ... line on stdout. The tool's own errors are one error: ... line on stderr. Both exit with the code shown.

From the app

From the app, for an action

From the tool

In apps, an app that couldn't be asked shows why in its note column instead, such as no reply in 5 s: JS may be paused at a breakpoint.

Exit codes

Timing

What it reads

Only what the stores keep, after every redaction in the config. The authorization, cookie and set-cookie headers and a page's document.cookie are always [redacted], in every output, and filters run on the masked row. An action's reply goes through the same masking. The only changes the tool can make are the actions. See Redaction and Older React Native.