🇺🇸English

next-browser

If next-browser is not already on PATH, install @vercel/next-browser globally with the user's package manager, then playwright install chromium.

If next-browser is already installed, it may be outdated. Run next-browser --version and compare against the latest on npm (npm view @vercel/next-browser version). If the installed version is behind, upgrade it (npm install -g @vercel/next-browser@latest or the equivalent for the user's package manager) before proceeding.

Next.js docs awareness

If the project's Next.js version is v16.2.0-canary.37 or later , bundled docs live at node_modules/next/dist/docs/. Before doing PPR work, Cache Components work, or any non-trivial Next.js task, read the relevant doc there — your training data may be outdated. The bundled docs are the source of truth.

See https://nextjs.org/docs/app/guides/ai-agents for background.

Working with the user

Onboarding

• If the user already gave a URL, cookies, and task — skip questions, open and go.
• Otherwise ask only what's missing: dev server URL (running?), session cookies if behind login.
• For cookies, give the user two options: (1) DevTools → Application → Cookies, export as [{"name":"session","value":"..."}], or (2) just "Copy as cURL" from DevTools → Network on any authenticated request — you can extract the cookies from the header yourself.
• Never say "ready, what would you like to do?". Never auto-discover (port scans, project, config reads) before being asked.

Show, don't tell

• screenshot after every navigation, code change, or visual finding. Always caption it (screenshot "Before fix", screenshot "PPR shell — locked"). In headed mode the Screenshot Log window opens automatically so the user sees every screenshot in real time.
• Don't narrate what a screenshot shows. State your conclusion or next action.

Escalate, don't decide

• Suspense boundary placement and fallback UI — design with the user.
• Caching decisions (staleness, visibility) — the user's call, not yours.
• "Make this page faster" without context — ask: cold URL hit or client navigation? From which page? Don't guess, don't do both.

Headless mode

By default the browser opens headed (visible window). For CI or cloud environments with no display, set NEXT_BROWSER_HEADLESS=1 to run headless.

Commands

open <url> [--cookies-json <file>]

Launch browser, navigate to URL. With --cookies-json, sets auth cookies before navigating (domain derived from URL hostname).

$ next-browser open http://localhost:3024/vercel --cookies-json cookies.json
opened → http://localhost:3024/vercel (11 cookies for localhost)

Cookie file format: [{"name":"authorization","value":"Bearer ..."}, ...]

Only name and value are required per cookie — omit domain, path, expires, etc. To create the file, use Bash (echo '[...]' > /tmp/cookies.json) since the Write tool requires a prior Read.

close

Close browser and kill daemon.

goto <url>

Navigate to a URL with a fresh server render. The browser loads a new document — equivalent to typing a URL in the address bar.

$ next-browser goto http://localhost:3024/vercel/~/deployments
→ http://localhost:3024/vercel/~/deployments

push [path]

Client-side navigation — the page transitions without a full reload, the way a user clicks a link in the app. Without a path, shows an interactive picker of all links on the current page.

$ next-browser push /vercel/~/deployments
→ http://localhost:3024/vercel/~/deployments

If push fails silently (URL unchanged), the route wasn't prefetched.

back

Go back one page in browser history.

reload

Reload the current page from the server.

ssr lock

Block external scripts on all subsequent navigations. While locked, every goto, push, back, and reload shows the raw server-rendered HTML without React hydration or client-side JavaScript — what search engines and social crawlers see.

$ next-browser ssr lock
ssr locked — external scripts blocked on all navigations

ssr unlock

Re-enable external scripts. The next navigation will load normally with full hydration.

$ next-browser ssr unlock
ssr unlocked — external scripts re-enabled

perf [url]

Profile a full page load — reloads the current page (or navigates to a URL) and collects Core Web Vitals and React hydration timing in one pass.

$ next-browser perf http://localhost:3000/dashboard
# Page Load Profile — http://localhost:3000/dashboard

## Core Web Vitals
  TTFB                   42ms
  LCP               1205.3ms (img: /_next/image?url=...)
  CLS                    0.03

## React Hydration — 65.5ms (466.2ms → 531.7ms)
  Hydrated                         65.5ms  (466.2 → 531.7)
  Commit                            2.0ms  (531.7 → 533.7)
  Waiting for Paint                 3.0ms  (533.7 → 536.7)
  Remaining Effects                 4.1ms  (536.7 → 540.8)

## Hydrated components (42 total, sorted by duration)
  DeploymentsProvider                       8.3ms
  NavigationProvider                        5.1ms
  ...

TTFB — server response time (Navigation Timing API). LCP — when the largest visible element painted, plus what it was. CLS — cumulative layout shift score (lower is better). Hydration — React reconciler phases and per-component cost (requires React profiling build / next dev; production strips console.timeStamp).

Without a URL, reloads the current page. With a URL, navigates there first.

renders start

Begin recording React re-renders. Hooks into onCommitFiberRoot to collect raw per-component data: render count, totalTime, selfTime, DOM mutations, change reasons, and FPS.

Survives full-page navigations (goto/reload) and captures mount and hydration renders — no need to start before or after navigation.

$ next-browser renders start
recording renders — interact with the page, then run `renders stop`

renders stop [--json]

Stop recording and print a per-component render profile. Raw data — the agent decides what's actionable.

$ next-browser renders stop
# Render Profile — 3.05s recording
# 426 renders (38 mounts + 388 re-renders) across 38 components
# FPS: avg 120, min 106, max 137, drops (<30fps): 0

## Components by total render time
| Component              | Insts | Mounts | Re-renders | Total    | Self     | DOM   | Top change reason          |
| ---------------------- | ----- | ------ | ---------- | -------- | -------- | ----- | -------------------------- |
| Parent                 |     1 |      1 |          9 |   5.8ms  |   3.4ms  | 10/10 | state (hook #0)            |
| MemoChild              |     3 |      3 |         27 |     2ms  |   1.9ms  | 30/30 | props.data                 |
| Router                 |     1 |      1 |          9 |   6.3ms  |       —  |  0/10 | parent (ErrorBoundaryHandler) |

## Change details (prev → next)
  Parent
    state (hook #0): 0 → 1
  MemoChild
    props.data: {value} → {value}

The Change details section shows the actual prev→next values for each change. This makes the data self-contained — you can see that MemoChild gets props.data: {value} → {value} (same shape, new reference — memo defeated) without needing to inspect the component.

With--json, outputs raw structured JSON with full change arrays per component (type, name, prev, next for each render event).

Columns:

• Insts — number of unique component instances observed during recording
• Mounts — how many times an instance mounted (first render, no alternate fiber)
• Re-renders — update-phase renders (total renders minus mounts)
• Total — inclusive render time (component + children)
• Self — exclusive render time (component only, excludes children)
• DOM — how many renders actually mutated the DOM vs total renders
• Top change reason — most frequent trigger for this component

Timing data (Total, Self) requires a React profiling build (next dev). In production builds these columns show — but render counts, DOM mutations, and change reasons are still reported.

Change reasons — what triggered each re-render:

• props.<name> — a prop changed by reference, with prev→next values
• state (hook #N) — a useState/useReducer hook changed, with prev→next values
• context (<name>) — a specific context changed, with prev→next values
• parent (<name>) — parent component re-rendered, names the parent
• parent (<name> (mount)) — parent is also mounting (typical during page load, not a leak)
• mount — first render

FPS — frames per second during recording. drops counts frames below 30fps.

Up to 200 components are tracked. If output exceeds 4 000 chars it is written to a temp file.

restart-server

Restart the Next.js dev server and clear its caches. Forces a clean recompile from scratch.

Last resort. HMR picks up code changes on its own — reach for this only when you have evidence the dev server is wedged (stale output after edits, builds that never finish, errors that don't clear).

Often exits with net::ERR_ABORTED — this is expected (the page detaches during restart). Follow up with goto <url> to re-navigate after the server is back. Don't treat this error as a failure.

ppr lock

Prerequisite: PPR requires cacheComponents to be enabled in next.config. Without it the shell won't have pre-rendered content to show.

Freeze dynamic content so you can inspect the static shell — the part of the page that's instantly available before any data loads. After locking:

• goto — shows the server-rendered shell with holes where dynamic content would appear.

• push — shows what the client already has from prefetching. Requires the current page to already be hydrated (prefetch is client-side), so lock after you've landed on the origin, not before.

  $ next-browser ppr lock locked

ppr unlock

Resume dynamic content and print a shell analysis — which Suspense boundaries were holes in the shell, what blocked them, and which were static. The output can be very large (hundreds of boundaries). Pipe through | head -20 if you only need the summary and dynamic holes.

$ next-browser ppr unlock
unlocked

# PPR Shell Analysis
# 131 boundaries: 3 dynamic holes, 128 static

## Summary
- Top actionable hole: TrackedSuspense — usePathname (client-hook)
- Suggested next step: This route segment is suspending on client hooks. Check loading.tsx first...
- Most common root cause: usePathname (client-hook) affecting 1 boundary

## Quick Reference
| Boundary                   | Type              | Fallback source | Primary blocker           | Source                        | Suggested next step       |
| ---                        | ---               | ---             | ---                       | ---                           | ---                       |
| TrackedSuspense            | component         | unknown         | usePathname (client-hook) | tracked-suspense.js:6         | Push the hook-using cl... |
| TeamDeploymentsLayout      | route-segment     | unknown         | unknown                   | layout.tsx:37                 | Inspect the nearest us... |
| Next.Metadata              | component         | unknown         | unknown                   | unknown                       | No primary blocker was... |

## Detail
  TrackedSuspense
    rendered by: TrackedSuspense > RootLayout > AppLayout
    environments: SSR
  TeamDeploymentsLayout
    suspenders unknown: thrown Promise (library using throw instead of use())

## Static (pre-rendered in shell)
  GeistProvider at .../geist-provider.tsx:80:9
  TrackedSuspense at ...
  ...

The Quick Reference table is the main overview — boundary, blocker, source, and suggested fix at a glance. The Detail section only appears for holes that have extra info (owner chains, environments, secondary blockers) not already in the table.

errors doesn't report while locked. If the shell looks wrong (empty, bailed to CSR), unlock and goto the page normally, then run errors. Don't debug blind under the lock.

Full bailout (scrollHeight = 0). When PPR bails out completely, unlock returns just "unlocked" with no shell analysis — there are no boundaries to report. In this case, unlock, goto the page normally, then use errors and logs to find the root cause.

tree

Full React component tree — every component on the page with its hierarchy, like the Components panel in React DevTools.

$ next-browser tree
# React component tree
# Columns: depth id parent name [key=...]
# Use `tree <id>` for props/hooks/state. IDs valid until next navigation.

0 38167 - Root
1 38168 38167 HeadManagerContext.Provider
2 38169 38168 Root
...
224 46375 46374 DeploymentsProvider
226 46506 46376 DeploymentsTable

tree <id>

Inspect one component: ancestor path, props, hooks, state, source location (source-mapped to original file).

$ next-browser tree 46375
path: Root > ... > Prerender(TeamDeploymentsPage) > Prerender(FullHeading) > Prerender(TrackedSuspense) > Suspense > DeploymentsProvider
DeploymentsProvider #46375
props:
  children: [<Lazy />, <Lazy />, <span />, <Lazy />, <Lazy />]
hooks:
  IsMobile: undefined (1 sub)
  Router: undefined (2 sub)
  DeploymentListScope: undefined (1 sub)
  User: undefined (4 sub)
  Team: undefined (4 sub)
  ...
  DeploymentsInfinite: undefined (12 sub)
source: app/(dashboard)/[teamSlug]/(team)/~/deployments/_parts/context.tsx:180:10

IDs are valid until navigation. Re-run tree after goto/push.

viewport [WxH]

Show or set the browser viewport size. Useful for testing responsive layouts.

$ next-browser viewport
1440x900

$ next-browser viewport 375x812
viewport set to 375x812

Once set, the viewport stays fixed across navigations. window.resizeTo() via eval is a no-op in Playwright — always use this command to change dimensions.

screenshot [caption] [--full-page]

Behavioral rules are in Working with the user → Show, don't tell.

Use screenshot only when visual layout matters (CSS, appearance, PPR shell). For page content or deciding what to click, use snapshot.

Captures the viewport (or full scrollable page with --full-page) to a temp PNG file and returns the path. In headed mode, every screenshot is added to the Screenshot Log — a live browser window that accumulates all screenshots taken during the session. In headless mode the log window is skipped.

The optional caption describes the screenshot or the rationale for taking it. Captions appear in the Screenshot Log above each image.

$ next-browser screenshot "Homepage after login"
/tmp/next-browser-1711234567890.png

$ next-browser screenshot "Full page layout" --full-page
/tmp/next-browser-1711234567891.png

snapshot

Snapshot the page's accessibility tree — the semantic structure a screen reader sees — with [ref=eN] markers on every interactive element. Use this to discover what's on the page before clicking.

$ next-browser snapshot
- navigation "Main"
  - link "Home" [ref=e0]
  - link "Dashboard" [ref=e1]
- main
  - heading "Settings"
  - tablist
    - tab "General" [ref=e2] (selected)
    - tab "Security" [ref=e3]
  - region "Profile"
    - textbox "Username" [ref=e4]
    - button "Save" [ref=e5]

The tree shows headings, landmarks (navigation, main, region), and state (selected, checked, expanded, disabled) so you understand page layout, not just a flat element list.

Refs are ephemeral — they reset on every snapshot call and are invalid after navigation. Re-run snapshot after goto/push.

click <ref|text|selector>

Click an element using real pointer events (pointerdown → mousedown → pointerup → mouseup → click). This works with libraries that ignore synthetic .click() (Radix UI, Headless UI, etc.).

Three ways to target:

Recommended workflow: run snapshot first, then click eN. Refs are the most reliable — they resolve via ARIA role+name, so they work even when elements have no stable CSS selector.

Clicking navigation links can timeout. click on a Next.js <Link> waits for the navigation to settle, which can exceed the command timeout. If click hangs on a nav link, cancel it and use goto <url> instead.

$ next-browser snapshot
- tablist
  - tab "General" [ref=e0] (selected)
  - tab "Security" [ref=e1]
$ next-browser click e1
clicked
$ next-browser snapshot
- tablist
  - tab "General" [ref=e0]
  - tab "Security" [ref=e1] (selected)

fill <ref|selector> <value>

Fill a text input or textarea. Clears existing content, then types the new value — dispatches all the events React and other frameworks expect.

$ next-browser snapshot
- textbox "Username" [ref=e4]
$ next-browser fill e4 "judegao"
filled

eval [ref] <script> · eval [ref] --file <path> · eval -

Run JS in page context. Returns the result as JSON.

With a ref , the script receives the DOM element as its argument — useful for inspecting a snapshot node or bridging to React internals:

$ next-browser eval e0 'el => el.tagName'
"BUTTON"

$ next-browser eval e0 'el => {
  const key = Object.keys(el).find(k => k.startsWith("__reactFiber$"));
  if (!key) return null;
  let fiber = el[key];
  while (fiber && typeof fiber.type !== "function") fiber = fiber.return;
  return fiber?.type?.displayName || fiber?.type?.name || null;
}'
"LoginButton"

For simple one-liners (no ref), pass the script inline:

$ next-browser eval 'document.title'
"Deployments – Vercel"

$ next-browser eval 'document.querySelectorAll("a[href]").length'
47

For multi-line or quote-heavy scripts , use --file (or -f) to avoid shell quoting issues entirely:

cat > /tmp/nb-eval.js << 'SCRIPT'
(() => {
  // your JS here — no shell escaping needed
  return someResult;
})()
SCRIPT
next-browser eval --file /tmp/nb-eval.js

You can also pipe via stdin: echo 'document.title' | next-browser eval -

Use this to read the Next.js error overlay (it's in shadow DOM): next-browser eval 'document.querySelector("nextjs-portal")?.shadowRoot?.querySelector("[data-nextjs-dialog]")?.textContent'

eval runs synchronously in page context — top-level await is not supported. Wrap in an async IIFE if you need to await: next-browser eval '(async () => { ... })()'.

errors

Build and runtime errors for the current page.

$ next-browser errors
{
  "configErrors": [],
  "sessionErrors": [
    {
      "url": "/vercel/~/deployments",
      "buildError": null,
      "runtimeErrors": [
        {
          "type": "console",
          "errorName": "Error",
          "message": "Route \"/[teamSlug]/~/deployments\": Uncached data or `connection()` was accessed outside of `<Suspense>`...",
          "stack": [
            {"file": "app/(dashboard)/.../deployments.tsx", "methodName": "Deployments", "line": 105, "column": 27}
          ]
        }
      ]
    }
  ]
}

buildError is a compile failure. runtimeErrors has type: "runtime" (React errors) and type: "console" (console.error calls).

logs

Recent dev server log output.

$ next-browser logs
{"timestamp":"00:01:55.381","source":"Server","level":"WARN","message":"[browser] navigation-metrics: skeleton visible was already recorded..."}
{"timestamp":"00:01:55.382","source":"Browser","level":"WARN","message":"navigation-metrics: content visible was already recorded..."}

browser-logs

Browser-side console output (console.log, console.warn, console.error, console.info). Captured directly from the page — works with both dev and production builds.

$ next-browser browser-logs
[LOG  ] Initializing app
[WARN ] Deprecation: use fetchV2 instead
[ERROR] Failed to load resource: 404
[INFO ] render complete in 42ms

Up to 500 entries are kept; oldest are dropped when the buffer is full. Entries accumulate across navigations within the same browser session. If output exceeds 4 000 chars it is written to a temp file and the path is printed instead.

When to use which:

For dev server diagnostics, prefer logs and errors. Use browser-logs when you need general console output or are running a production build.

network

List all network requests since last navigation.

$ next-browser network
# Network requests since last navigation
# Columns: idx status method type ms url [next-action=...]
# Use `network <idx>` for headers and body.

0 200 GET document 508ms http://localhost:3024/vercel
1 200 GET font 0ms http://localhost:3024/_next/static/media/797e433ab948586e.p.d2077940.woff2
2 200 GET stylesheet 6ms http://localhost:3024/_next/static/chunks/_a17e2099._.css
3 200 GET fetch 102ms http://localhost:3024/api/v9/projects next-action=abc123def

Server actions show next-action=<id> suffix.

network <idx>

Full request/response for one entry. Long bodies spill to temp files.

$ next-browser network 0
GET http://localhost:3024/vercel
type: document  508ms

request headers:
  accept: text/html,...
  cookie: authorization=Bearer...; isLoggedIn=1; ...
  user-agent: Mozilla/5.0 ...

response: 200 OK
response headers:
  cache-control: no-cache, must-revalidate
  content-encoding: gzip
  ...

response body:
(8234 bytes written to /tmp/next-browser-12345-0.html)

page

Route segments for the current URL — which layouts, pages, and boundaries are active.

$ next-browser page
{
  "sessions": [
    {
      "url": "/vercel/~/deployments",
      "routerType": "app",
      "segments": [
        {"path": "app/(dashboard)/[teamSlug]/(team)/~/deployments/layout.tsx", "type": "layout", ...},
        {"path": "app/(dashboard)/[teamSlug]/(team)/~/deployments/page.tsx", "type": "page", ...},
        {"path": "app/(dashboard)/[teamSlug]/layout.tsx", "type": "layout", ...},
        {"path": "app/(dashboard)/layout.tsx", "type": "layout", ...},
        {"path": "app/layout.tsx", "type": "layout", ...}
      ]
    }
  ]
}

project

Project root and dev server URL.

$ next-browser project
{
  "projectPath": "/Users/judegao/workspace/repo/front/apps/vercel-site",
  "devServerUrl": "http://localhost:3331"
}

routes

All app router routes.

$ next-browser routes
{
  "appRouter": [
    "/[teamSlug]",
    "/[teamSlug]/~/deployments",
    "/[teamSlug]/[project]",
    "/[teamSlug]/[project]/[id]/logs",
    ...
  ]
}

action <id>

Inspect a server action by its ID (from next-action header in network list).

Scenarios

Debugging rendering performance

When the user says "this page is slow after load", "too many re-renders", "laggy interactions", or "janky" — this is update-phase rendering, not initial load. Use renders to profile it. (For initial load, use perf.)

Workflow:

1. renders start — begin recording.
2. goto the page (the hook survives navigation and captures mount).
3. Reproduce the slow interaction: click buttons, type in inputs, navigate via push, or just wait if the issue is polling/timers.
4. renders stop — read the raw data.
5. Use the data to form hypotheses. The columns give you:
   • Mounts vs Re-renders — is this component re-rendering after load, or is the count just from mount-time cascading?
   • Insts — is a high render count from many instances or one instance rendering excessively?
   • Self — is this component expensive per-render, or just called too often?
   • DOM — did the renders actually produce visible changes? A component with 100 renders and 0 DOM mutations is doing purely wasted work.

Verify the fix. After editing the code, HMR picks it up. Re-run renders start / renders stop and compare the raw numbers to the previous profile.

Test your hypothesis before proposing a fix. If you suspect a component is the root cause, find evidence — inspect it with tree, read its source, check what's changing via the change reason column. Don't propose changes from a single observation.

Growing the static shell

The shell is what the user sees the instant they land — before any dynamic data arrives. The measure is the screenshot while locked: does it read as the page itself? A shell can be non-empty and still bad — one Suspense fallback wrapping the whole content area renders something , but it's a monolithic loading state, not the page.

A meaningful shell is the real component tree with small, local fallbacks where data is genuinely pending. Getting there means the composition layer — the layouts and wrappers between those leaf boundaries — can't itself suspend. ppr unlock's Quick Reference table names the primary blocker and source for each hole; the Detail section adds owner chains and secondary blockers. A suspend high in the tree is what collapses everything beneath it into one fallback.

Work it top-down. For the component that's suspending: can the dynamic access move into a child? If yes, move it — this component becomes sync and rejoins the shell. Follow the access down and ask again.

When you reach a component where it can't move any lower, there are two exits — wrap in a Suspense boundary, or cache it for prerender. Both are human calls (see Working with the user → Escalate, don't decide).

Test your hypothesis before proposing a fix. If you suspect a component is the cause, find evidence — check errors, inspect the component with tree, or compare a route where the shell works to one where it doesn't. Don't commit to a root cause or propose changes from a single observation.

There are two shells depending on how the user arrives — establish which one you're optimizing first (see Working with the user → Escalate, don't decide).

Direct load — the PPR shell. Server HTML for a cold hit on the URL. Lock first, then goto the target — the lock suppresses hydration so you see exactly what the server sent. Screenshot once the load settles, then unlock.

Client navigation — the prefetched shell. What the router already holds when a link is clicked. The origin page decides this — it's the one doing the prefetching — so goto the origin unlocked and let it fully hydrate. Then lock, push to the target, let the navigation settle, screenshot, unlock. Locking before the origin hydrates means nothing got prefetched and push has nothing to show.

Between iterations: check errors while unlocked.

After making a code change: HMR picks it up — just re-lock, goto the page, and re-test. No need to restart-server.

Weekly Installs

1.1K

Repository

vercel-labs/next-browser

GitHub Stars

83

First Seen

Mar 6, 2026

Security Audits

Gen Agent Trust HubPassSocketPassSnykFail

Installed on

opencode1.0K

codex1.0K

github-copilot1.0K

gemini-cli1.0K

cursor1.0K

cline1.0K