openclaw/docs/tools/browser-control.md

---
summary: "OpenClaw browser control API, CLI reference, and scripting actions"
read_when:
  - Scripting or debugging the agent browser via the local control API
  - Looking for the `openclaw browser` CLI reference
  - Adding custom browser automation with snapshots and refs
title: "Browser control API"
---

For setup, configuration, and troubleshooting, see [Browser](/tools/browser).
This page is the reference for the local control HTTP API, the `openclaw browser`
CLI, and scripting patterns (snapshots, refs, waits, debug flows).

## Control API (optional)

For local integrations only, the Gateway exposes a small loopback HTTP API.
This standalone server is opt-in — set the environment variable
`OPENCLAW_EAGER_BROWSER_CONTROL_SERVER=1` in the gateway service environment
and restart the gateway before the HTTP endpoints become available. Without
this variable the browser control runtime still works through the CLI and
agent tools, but nothing listens on the loopback control port.

- Status/start/stop: `GET /`, `GET /doctor`, `POST /start`, `POST /stop`, `POST /reset-profile`
- Profiles: `GET /profiles`, `POST /profiles/create`, `DELETE /profiles/:name`
- Tabs: `GET /tabs`, `POST /tabs/open`, `POST /tabs/focus`, `DELETE /tabs/:targetId`, `POST /tabs/action`
- Snapshot/screenshot/stream: `GET /snapshot`, `POST /screenshot`, `POST /screencast`
- Actions: `POST /navigate`, `POST /act`
- Hooks: `POST /hooks/file-chooser`, `POST /hooks/dialog`
- Downloads: `POST /download`, `POST /wait/download`
- Permissions: `POST /permissions/grant`
- Debugging: `GET /console`, `GET /errors`, `GET /requests`, `GET /dialogs`, `POST /pdf`, `POST /trace/start`, `POST /trace/stop`, `POST /highlight`
- Network: `POST /response/body`
- State: `GET /cookies`, `POST /cookies/set`, `POST /cookies/clear`, `GET /storage/:kind`, `POST /storage/:kind/set`, `POST /storage/:kind/clear`
- Settings: `POST /set/offline`, `POST /set/headers`, `POST /set/credentials`, `POST /set/geolocation`, `POST /set/media`, `POST /set/timezone`, `POST /set/locale`, `POST /set/device`

`POST /tabs/action` is the batched form the CLI uses internally for
`browser tab` subcommands (`{"action":"new"|"label"|"select"|"close"|"list", ...}`).
Prefer the single-purpose tab routes above when scripting directly.

All endpoints accept `?profile=<name>`. `POST /start?headless=true` requests a
one-shot headless launch for local managed profiles without changing persisted
browser config. Attach-only, remote CDP, and existing-session profiles reject
that override because OpenClaw does not launch those browser processes.

For tab endpoints, `targetId` is the compatibility field name. Prefer passing
`suggestedTargetId` from `GET /tabs` or `POST /tabs/open`. Labels and `tabId`
handles such as `t1` are also accepted. Raw CDP target ids and unique raw
target-id prefixes still work, but they are volatile diagnostic handles.
Tab handles are scoped to a browser host or node and profile. Keep that route
with the handle when making follow-up requests.

For profiles configured with `driver: "extension"`, `GET /tabs` and the browser
tool can also return `webExtensionTabId`, the runtime-scoped numeric Chrome
WebExtensions tab ID for the same tab. This field is omitted for other drivers
or when extension metadata is unavailable. Use it only when calling a
WebExtensions API; continue to use `suggestedTargetId` or `tabId` for OpenClaw
browser actions because a `webExtensionTabId` can change after the browser or
extension reconnects.

The Control UI's `browser.request` Gateway method accepts `target: "host"` to
pin the Gateway host or `target: "node"` with `node: "<node-id>"` to pin a browser
node. Pass the profile in `query.profile`. Explicit routes do not fall back to
another host. Omitting them keeps the configured automatic routing. These
routing fields do not grant access or change browser policy.

Browser previews require a result from the `browser` tool with a known route.
Browser-shaped metadata from other tools does not trigger screenshots or change
the panel's selection. Those results remain ordinary tool output.

When URL validation fails during tab listing, the tab keeps its identity and
title but returns `url: ""` and `urlUnavailableReason`:

- `navigation_blocked`: navigation rules rejected the address.
- `navigation_check_failed`: OpenClaw could not validate the address, for example
  because DNS lookup failed. Refresh to check again.

An empty URL alone does not indicate a policy denial. Navigation-policy errors
also carry `reason: "navigation_blocked"`. Raw blocked URLs and DNS details are
not included in that metadata. Tab listings are observations, not authorization:
every subsequent content read or action still enforces its own checks.

If shared-secret gateway auth is configured, browser HTTP routes require auth too:

- `Authorization: Bearer <gateway token>`
- `x-openclaw-password: <gateway password>` or HTTP Basic auth with that password

Notes:

- This standalone loopback browser API does **not** consume trusted-proxy or
  Tailscale Serve identity headers.
- If `gateway.auth.mode` is `none` or `trusted-proxy`, these loopback browser
  routes do not inherit those identity-bearing modes. Keep them loopback-only.

### Screencast stream

`POST /screencast` mints a single-use token for a live view of the selected tab.
Pass an optional `targetId`, `maxWidth`, `maxHeight`, and `quality` in the JSON
body. Dimensions default to 1280 and are clamped to integers from 320 to 2000.
JPEG quality defaults to 70 and is clamped from 30 to 90.

The response contains `token`, `wsPath`, `expiresAtMs`, `targetId`, and `url`.
Resolve `wsPath` against the Gateway URL and open a WebSocket there:
`/browser/screencast?token=<token>`. The 48-character hexadecimal token expires
after 60 seconds and can be consumed once. Invalid, expired, and reused tokens
are rejected with HTTP 401 before upgrade. Viewers send no application messages.
Binary viewer messages close the connection.

Tickets and viewers minted through the Gateway are bound to the requesting
Gateway connection. Ending that connection revokes unused tickets and closes
its viewers. Revoked (invalidated) connections are fenced immediately, before
the Gateway socket finishes closing: their tickets cannot upgrade, and their
viewers receive no further frames or metadata. The loopback HTTP control API
has no Gateway connection to bind, so its tickets remain TTL-only.

The plugin shares one CDP screencast per profile and tab. Chrome sends JPEG
frames on repaint, paced to approximately 20 frames per second. Slow viewers
skip frames instead of building a queue. Navigation immediately retires the
capture session. A new CDP session starts only after the address is allowed,
so delayed frames from the previous document cannot enter the new stream.
A rejected navigation stops the stream.

Text messages are JSON with `type` in `ready`, `meta`, or `error`. A `ready`
message includes `targetId`, `url`, and `title`. `meta` updates `url` and `title`
after allowed navigation and page load. Binary messages contain:

1. A four-byte unsigned big-endian JSON header length.
2. That many bytes of UTF-8 JSON: `{ "url", "cssWidth", "cssHeight", "scrollX", "scrollY", "ts" }`.
3. The JPEG bytes.

`cssWidth` and `cssHeight` come from CDP's `deviceWidth` and `deviceHeight`
metadata and describe the layout viewport in CSS pixels. `scrollX` and `scrollY`
come from `scrollOffsetX` and `scrollOffsetY`. `ts` is the CDP frame timestamp.

| Close code | Meaning                                                                                                     |
| ---------- | ----------------------------------------------------------------------------------------------------------- |
| 4001       | Token invalid or expired (normally rejected before upgrade with HTTP 401)                                   |
| 4003       | `navigation_blocked`                                                                                        |
| 4004       | `target_closed`, including a profile lifecycle change                                                       |
| 4005       | Unsupported streaming                                                                                       |
| 4006       | `authority_revoked` (the requesting Gateway connection ended or was invalidated, e.g. device token revoked) |
| 1012       | Gateway shutting down                                                                                       |

Chrome MCP existing-session profiles and missing Playwright return HTTP 501 with
`code: "SCREENCAST_UNSUPPORTED"` and `reason: "existing-session"` or `"playwright"`.
Node-routed requests fail before proxying with `INVALID_REQUEST` and details
`{ "code": "SCREENCAST_UNSUPPORTED", "reason": "node" }`. The Control UI falls
back to the existing screenshot route when streaming is unavailable.
Navigation metadata updates the tab and address bar. The displayed image keeps
its own URL and metrics until a replacement frame arrives. While Annotate or
Inspect is active, the Control UI pins the captured image and its URL, then
displays the latest held frame when capture mode ends.

### `/act` error contract

`POST /act` uses a structured error response for validation, policy, and
recognized interaction failures:

```json
{ "error": "<message>", "code": "ACT_*" }
```

Current `code` values:

- `ACT_KIND_REQUIRED` (HTTP 400): `kind` is missing or unrecognized.
- `ACT_INVALID_REQUEST` (HTTP 400): action payload failed normalization or validation.
- `ACT_SELECTOR_UNSUPPORTED` (HTTP 400): `selector` was used with an unsupported action kind.
- `ACT_EVALUATE_DISABLED` (HTTP 403): `evaluate` (or `wait --fn`) is disabled by config.
- `ACT_TARGET_ID_MISMATCH` (HTTP 403): top-level or batched `targetId` conflicts with request target.
- `ACT_OPERATION_FAILED` (HTTP 500): the selected element could not perform the action, such as a noneditable input, covered control, or ambiguous ref. The message describes the interaction failure without treating it as a browser connection outage.
- `ACT_EXISTING_SESSION_UNSUPPORTED` (HTTP 501): action is not supported for existing-session profiles.

Other runtime failures may still return `{ "error": "<message>" }` without a
`code` field.

### Playwright requirement

Some features (navigate/act/AI snapshot/role snapshot, element
screenshots, PDF) require Playwright. If Playwright isn't installed, those endpoints return
a clear 501 error.

What still works without Playwright:

- ARIA snapshots
- Role-style accessibility snapshots (`--interactive`, `--compact`,
  `--depth`, `--efficient`) when a per-tab CDP WebSocket is available. This is
  a fallback for inspection and ref discovery. Playwright remains the primary
  action engine.
- Page screenshots for the managed `openclaw` browser when a per-tab CDP
  WebSocket is available
- Page screenshots for `existing-session` / Chrome MCP profiles
- `existing-session` ref-based screenshots (`--ref`) from snapshot output

What still needs Playwright:

- `navigate`
- `act`
- AI snapshots that depend on Playwright's native AI snapshot format
- CSS-selector element screenshots (`--element`)
- full browser PDF export

Element screenshots also reject `--full-page`. The route returns `fullPage is
not supported for element screenshots`.

If you see `Playwright is not available in this gateway build`, the packaged
Gateway is missing the core browser runtime dependency. Reinstall or update
OpenClaw, then restart the gateway. For Docker, also install the Chromium
browser binaries as shown below.

#### Docker Playwright install

If your Gateway runs in Docker, avoid `npx playwright` (npm override conflicts).
For custom images, bake Chromium into the image:

```bash
OPENCLAW_INSTALL_BROWSER=1 ./scripts/docker/setup.sh
```

The browser also needs system libraries, so installing Chromium in a one-off
Compose container is not durable. Rebuild the image with
`OPENCLAW_INSTALL_BROWSER=1` instead. To persist browser downloads and other
caches, persist `/home/node` with `OPENCLAW_HOME_VOLUME` or a bind mount. See
[Docker](/install/docker).

## How it works (internal)

A small loopback control server accepts HTTP requests and connects to Chromium-based browsers via CDP. Advanced actions (click/type/snapshot/PDF) go through Playwright on top of CDP. When Playwright is missing, only non-Playwright operations are available. The agent sees one stable interface while local/remote browsers and profiles swap freely underneath.

## CLI quick reference

All commands accept `--browser-profile <name>` to target a specific profile, and `--json` for machine-readable output.

<AccordionGroup>

<Accordion title="Basics: status, tabs, open/focus/close">

```bash
openclaw browser status
openclaw browser doctor
openclaw browser doctor --deep    # add a live snapshot probe
openclaw browser start
openclaw browser start --headless # one-shot local managed headless launch
openclaw browser stop            # also clears emulation on attach-only/remote CDP
openclaw browser reset-profile   # moves the profile's browser data to Trash
openclaw browser tabs
openclaw browser tab             # shortcut for current tab
openclaw browser tab new
openclaw browser tab new --label research
openclaw browser tab label abcd1234 research
openclaw browser tab select 2
openclaw browser tab close 2
openclaw browser open https://example.com
openclaw browser focus abcd1234
openclaw browser close abcd1234
```

</Accordion>

<Accordion title="Profiles: list, create, delete">

```bash
openclaw browser profiles
openclaw browser create-profile --name research --color "#0066CC"
openclaw browser create-profile --name attach --driver existing-session --cdp-url http://127.0.0.1:9222
openclaw browser delete-profile --name research
```

</Accordion>

<Accordion title="Inspection: screenshot, snapshot, console, errors, requests">

```bash
openclaw browser screenshot
openclaw browser screenshot --full-page
openclaw browser screenshot --ref 12        # or --ref e12
openclaw browser screenshot --labels
openclaw browser snapshot
openclaw browser snapshot --format aria --limit 200
openclaw browser snapshot --interactive --compact --depth 6
openclaw browser snapshot --efficient
openclaw browser snapshot --labels
openclaw browser snapshot --urls
openclaw browser snapshot --selector "#main" --interactive
openclaw browser snapshot --frame "iframe#main" --interactive
openclaw browser snapshot --out snapshot.txt
openclaw browser console --level error
openclaw browser errors --clear
openclaw browser requests --filter api --clear
openclaw browser pdf
openclaw browser responsebody "**/api" --max-chars 5000
```

</Accordion>

<Accordion title="Actions: navigate, click, type, drag, wait, evaluate">

```bash
openclaw browser navigate https://example.com
openclaw browser resize 1280 720
openclaw browser click 12 --double           # or e12 for role refs
openclaw browser click-coords 120 340        # viewport coordinates
openclaw browser type 23 "hello" --submit
openclaw browser press Enter
openclaw browser hover 44
openclaw browser scrollintoview e12
openclaw browser drag 10 11
openclaw browser select 9 OptionA OptionB
openclaw browser download e12 report.pdf
openclaw browser waitfordownload report.pdf
openclaw browser upload /tmp/openclaw/uploads/file.pdf
openclaw browser upload /tmp/openclaw/uploads/file.pdf --ref e12
openclaw browser upload media://inbound/file.pdf
openclaw browser fill --fields '[{"ref":"1","type":"text","value":"Ada"}]'
openclaw browser dialog --accept
openclaw browser dialog --dismiss --dialog-id d1
openclaw browser wait --text "Done"
openclaw browser wait "#main" --url "**/dash" --load networkidle --fn "window.ready===true"
openclaw browser evaluate --fn '(el) => el.textContent' --ref 7
openclaw browser evaluate --fn 'const title = document.title; return title;'
openclaw browser evaluate --timeout-ms 30000 --fn 'async () => { await window.ready; return true; }'
openclaw browser highlight e12
openclaw browser trace start
openclaw browser trace stop
```

</Accordion>

<Accordion title="State: cookies, storage, offline, headers, geo, device">

```bash
openclaw browser cookies
openclaw browser cookies set session abc123 --url "https://example.com"
openclaw browser cookies clear
openclaw browser storage local get
openclaw browser storage local set theme dark
openclaw browser storage session clear
openclaw browser set offline on
openclaw browser set headers --headers-json '{"X-Debug":"1"}'
openclaw browser set credentials user pass            # --clear to remove
openclaw browser set geo 37.7749 -122.4194 --origin "https://example.com"
openclaw browser set media dark
openclaw browser set timezone America/New_York
openclaw browser set locale en-US
openclaw browser set device "iPhone 14"
```

</Accordion>

</AccordionGroup>

Notes:

- The agent-facing `browser` tool exposes `action=download` (required `ref` and
  `path`) and `action=waitfordownload` (optional `path`). Both return the saved
  download URL, suggested filename, and guarded local path. Explicit download
  interception is available for managed Playwright profiles. Existing-session
  profiles return an unsupported-operation error.
- Prefer atomic chooser uploads: pass the trigger `--ref` with the upload so OpenClaw arms and clicks in one request. Paths-only `upload` remains supported when a later trigger is intentional. Use `--input-ref` or `--element` to set a file input directly. `dialog` is an arming call. Run it before the click/press that triggers the dialog. If an action opens a modal, the action response includes `blockedByDialog` and `browserState.dialogs.pending`. Pass that `dialogId` to respond directly. Dialogs handled outside OpenClaw appear under `browserState.dialogs.recent`.
- Cancelling a pending locator click, typing, or upload operation leaves other tabs connected. Upload waiters belong to the selected tab. A new upload on that tab replaces its previous waiter.
- `click`/`type`/etc require a `ref` from `snapshot` (for example, Playwright ref `f1e12`, role ref `e12`, or actionable ARIA ref `ax12`). Copy the returned ref unchanged, including any frame prefix. CSS selectors are intentionally not supported for actions. Use `click-coords` when the visible viewport position is the only reliable target.
- Download and trace paths are constrained to OpenClaw temp roots: `/tmp/openclaw{,/downloads}` (fallback: `${os.tmpdir()}/openclaw/...`).
- `upload` accepts files from the OpenClaw temp uploads root and
  OpenClaw-managed inbound media. Managed inbound media can be referenced as
  `media://inbound/<id>`, sandbox-relative `media/inbound/<id>`, or a resolved
  path inside the managed inbound media directory. Nested media refs,
  traversal, symlinks, hardlinks, and arbitrary local paths are still rejected.
- `upload` can also set file inputs directly via `--input-ref` or `--element`; these operations honor the upload timeout.
- Dialog prompt text preserves whitespace and empty strings exactly. Reading all local or session storage preserves empty keys and keys such as `__proto__`.

Stable tab ids and labels survive Chromium raw-target replacement when OpenClaw
can prove the replacement tab, such as a unique old/new pair for the same URL or
a single old tab becoming a single new tab after form submission. Ambiguous
duplicate-URL replacements receive fresh handles. Raw target ids are still
volatile. Prefer `suggestedTargetId` from `tabs` in scripts.

Snapshot flags at a glance:

- `--format ai` (default with Playwright): AI snapshot with native Playwright refs, including frame-qualified refs such as `f1e12`.
- `--format aria`: accessibility tree with `axN` refs. When Playwright is available, OpenClaw binds refs with backend DOM ids to the live page. Follow-up actions can then use them. Otherwise treat the output as inspection-only.
- `--efficient` (or `--mode efficient`): compact role snapshot preset. Set `browser.snapshotDefaults.mode: "efficient"` to make this the default (see [Gateway configuration](/gateway/config-browser-ui-desktop#browser)).
- `--interactive`, `--compact`, `--depth`, `--selector` force a role snapshot with `ref=e12` refs. `--frame "<iframe>"` scopes role snapshots to an iframe.
- Selector- and frame-scoped role refs bind to the captured DOM controls, including shadow DOM and external `aria-owns` members. Reordering controls does not retarget those refs. Removed controls or failed bindings require a new snapshot instead of matching another control by name.
- In scoped role snapshots, ignored nodes and unnamed generic wrappers are transparent before depth filtering. Depth counts the remaining role nodes from the selected root; wrapper lines and ref numbers can differ from older snapshots. State attributes, URL appendices, and output limits still apply.
- A selector-scoped snapshot is a point-in-time observation. If no element matches at request time, it returns an empty snapshot immediately. It does not wait for the snapshot timeout. Use `openclaw browser wait "<selector>"` when the page is expected to add the element later.
- `--selector` does not change the behavior of page-wide or frame-scoped transport failures. Those still use the configured snapshot timeout and diagnostics.
- With Playwright, `--labels` adds a screenshot with overlayed ref labels
  (prints `MEDIA:<path>`) plus an `annotations` array with each ref's bounding
  box. On `screenshot`, Playwright-backed labels work with `--full-page`,
  `--ref`, and `--element`. On `snapshot`, the accompanying screenshot remains
  viewport-only. Existing-session/chrome-mcp profiles render overlay labels on
  page screenshots but do not return `annotations` or use the Playwright
  full-page/ref/element projection helper. Without Playwright or chrome-mcp,
  labeled screenshots are not available.
- `--urls` appends discovered link destinations to AI snapshots.

## Snapshots and refs

OpenClaw supports three "snapshot" styles:

- **AI snapshot (native refs)**: `openclaw browser snapshot` (default, `--format ai`)
  - Output: a text snapshot with refs such as `f1e12` and matching `refs` metadata.
  - Actions: `openclaw browser click f1e12`, `openclaw browser type f1e23 "hello"` (use your snapshot's refs).
  - Internally, the ref is resolved via Playwright's `aria-ref`.

- **Role snapshot (role refs like `e12`)**: `openclaw browser snapshot --interactive` (or `--compact`, `--depth`, `--selector`, `--frame`)
  - Output: a role-based list/tree with `[ref=e12]` (and optional `[nth=1]`).
  - Actions: `openclaw browser click e12`, `openclaw browser highlight e12`.
  - Internally, the ref is resolved via `getByRole(...)` (plus `nth()` for duplicates).
  - Names containing quotes, backslashes, or YAML punctuation remain actionable. Use the ref rather than reconstructing a locator from the displayed name.
  - A missing displayed name can mean an empty accessible name or one above Playwright's 900 UTF-16-unit limit. Keep using the returned ref.
  - Add `--labels` to include a screenshot with overlayed `e12` labels. On
    Playwright-backed profiles this also returns per-ref bounding-box metadata
    (`annotations[]`). A labeled element screenshot preserves the ref and frame
    from the snapshot that produced it.
  - Add `--urls` when link text is ambiguous and the agent needs concrete
    navigation targets. With `--frame`, the URL appendix comes from that frame.

- **ARIA snapshot (ARIA refs like `ax12`)**: `openclaw browser snapshot --format aria`
  - Output: the accessibility tree as structured nodes.
  - Actions: `openclaw browser click ax12` works when the snapshot path can bind
    the ref through Playwright and Chrome backend DOM ids.
- If Playwright is unavailable, ARIA snapshots can still be useful for
  inspection, but refs may not be actionable. Re-snapshot with `--format ai`
  or `--interactive` when you need action refs.
- When the driver exposes stable document identity, consecutive AI and role
  snapshots for the same profile, tab, document, and option family append
  `[new]` to ref-bearing lines absent from the previous snapshot. Navigation
  starts a fresh unmarked baseline, including same-URL iframe reloads when the
  snapshot includes that frame. Existing-session snapshots omit deltas.
  The first snapshot establishes the baseline without markers. Later responses
  also expose `newElements`, and add a count footer when the value is nonzero.
  Structured `--format aria` snapshots with `axN` refs do not use delta markers.
- Contributors: the raw-CDP fallback path has a Docker proof lane, described in
  [Docker test suites](/reference/test/docker).

Ref behavior:

- Refs are **not stable across navigations**. If something fails, re-run `snapshot` and use a fresh ref.
- A batch stops after a committed main-frame navigation—including a same-URL
  reload—or after the page closes. Its `aborted` summary reports the action
  number and skipped count. Take a fresh snapshot before issuing dependent
  actions, or use separate act calls when navigation is expected.
- `/act` returns the current raw `targetId` after action-triggered replacement
  when it can prove the replacement tab. Keep using stable tab ids/labels for
  follow-up commands.
- A role snapshot taken with `--frame` scopes role refs to that iframe until the next role snapshot.
- Unknown or stale `axN` refs fail fast instead of falling through to
  Playwright's `aria-ref` selector. Run a fresh snapshot on the same tab when
  that happens.

## Browser batch CLI

`openclaw browser batch` runs an array of nested `/act` actions in one `/act`
call (the same `kind="batch"` runtime reached through the agent tool), so CLI
users and scripts can combine actions like `wait`, `click`, `type`, and
`evaluate` into a single replayable plan without per-action round trips. Each
entry in `actions[]` is a `BrowserActRequest` — the closed union the `/act`
route accepts (`click`, `clickCoords`, `type`, `press`, `hover`,
`scrollIntoView`, `drag`, `select`, `fill`, `resize`, `wait`, `evaluate`,
`close`, `batch`) — not arbitrary `openclaw browser` subcommands. `batch` is
not supported on `profile="user"` and other existing-session (chrome-mcp)
profiles. Send actions individually there.

- CLI: `openclaw browser batch --actions '<json>'`, `openclaw browser batch
--actions-file plan.json`, or `openclaw browser batch --actions-file -` to
  read the JSON array from stdin. `--continue` sets `stopOnError=false`. The
  default is to stop on first error. `--target-id` scopes the whole batch to
  one tab. `--actions-file` and stdin input are capped at 1,000,000 bytes.
  Split larger plans into multiple batch commands.
- Ref lifecycle: refs come from a `snapshot` run before the batch (snapshot is
  not a nested action). A nested action that changes page state — such as a
  `click` that triggers navigation, or an `evaluate` that mutates the DOM — can
  invalidate earlier refs for the rest of the batch. Put state-changing actions
  first, or split into a follow-up batch after re-snapshotting. Navigation and
  re-snapshotting happen outside the batch (`openclaw browser navigate` /
  `snapshot`), since `open`, `navigate`, and `snapshot` are not `/act` kinds.
- Target id conflicts: a nested action may omit `targetId` or repeat the
  request-level `targetId`. An explicit nested `targetId` that resolves to a
  different tab is rejected with `ACT_TARGET_ID_MISMATCH` before any action
  runs. Batched actions share the request's tab by design.
- Error summary: the response is `{ "results": [{ "ok": true }, { "ok": false,
"error": "<message>" }, ...] }`, one entry per action in order. When
  `stopOnError` is the default, the array ends at the first failure. With
  `--continue` it covers every action. Any failed entry makes the CLI exit
  nonzero. Pass `--json` to preserve the full ordered response for scripts.
- Nested batches occupy one parent result. If a child action fails, that result
  reports the first child error. Each batch applies its own `stopOnError`:
  continuing inside a nested batch does not make it successful or make its
  parent continue.

## Wait power-ups

You can wait on more than just time/text:

- Wait for URL (globs supported by Playwright):
  - `openclaw browser wait --url "**/dash"`
- Wait for load state:
  - `openclaw browser wait --load networkidle`
  - Supported on managed `openclaw` and raw/remote CDP profiles. Profiles using the `existing-session` driver (including the default `user` profile) reject `networkidle`. Use `--url`, `--text`, a selector, or `--fn` waits there.
- Wait for a JS predicate:
  - `openclaw browser wait --fn "window.ready===true"`
- Wait for a selector to become visible:
  - `openclaw browser wait "#main"`

These can be combined:

```bash
openclaw browser wait "#main" \
  --url "**/dash" \
  --load networkidle \
  --fn "window.ready===true" \
  --timeout-ms 15000
```

## Debug workflows

When an action fails (e.g. "not visible", "strict mode violation", "covered"):

1. `openclaw browser snapshot --interactive`
2. Use `click <ref>` / `type <ref>` (prefer role refs in interactive mode)
3. If it still fails: `openclaw browser highlight <ref>` to see what Playwright is targeting
4. If the page behaves oddly:
   - `openclaw browser errors --clear`
   - `openclaw browser requests --filter api --clear`
5. For deep debugging: record a trace:
   - `openclaw browser trace start`
   - reproduce the issue
   - `openclaw browser trace stop` (prints `TRACE:<path>`)

## JSON output

`--json` is for scripting and structured tooling.

Examples:

```bash
openclaw browser --json status
openclaw browser --json snapshot --interactive
openclaw browser --json requests --filter api
openclaw browser --json cookies
```

Role snapshots in JSON include `refs` plus a small `stats` block (lines/chars/refs/interactive) so tools can reason about payload size and density.

## State and environment knobs

These are useful for "make the site behave like X" workflows:

- Cookies: `cookies`, `cookies set`, `cookies clear`
- Storage: `storage local|session get|set|clear`
- Offline: `set offline on|off`
- Headers: `set headers --headers-json '{"X-Debug":"1"}'` (or the positional form `set headers '{"X-Debug":"1"}'`)
- HTTP basic auth: `set credentials user pass` (or `--clear`)
- Geolocation: `set geo <lat> <lon> --origin "https://example.com"` (or `--clear`)
- Media: `set media dark|light|no-preference|none`
- Timezone / locale: `set timezone ...`, `set locale ...`
- Device / viewport:
  - `set device "iPhone 14"` (Playwright device presets)
  - `set viewport 1280 720`

## Security and privacy

- The openclaw browser profile may contain logged-in sessions. Treat it as sensitive.
- `browser act kind=evaluate` / `openclaw browser evaluate` and `wait --fn`
  execute arbitrary JavaScript in the page context. Prompt injection can steer
  this. Disable it with `browser.evaluateEnabled=false` if you do not need it.
- `openclaw browser evaluate --fn` accepts a function source, an expression, or
  a statement body. Statement bodies are wrapped as async functions, so use
  `return` for the value you want back. Use `--timeout-ms <ms>` when the
  page-side function may need longer than the default evaluate timeout.
- For logins and anti-bot notes (X/Twitter, etc.), see [Browser login + X/Twitter posting](/tools/browser-login).
- Keep the Gateway/node host private (loopback or tailnet-only).
- Remote CDP endpoints are powerful. Tunnel and protect them.

Strict-mode example (block private/internal destinations by default):

```json5
{
  browser: {
    ssrfPolicy: {
      dangerouslyAllowPrivateNetwork: false,
      allowedHostnames: ["*.example.com", "example.com", "localhost"],
    },
  },
}
```

## Related

- [Browser](/tools/browser) - overview, configuration, profiles, security
- [Browser login](/tools/browser-login) - signing in to sites
- [Browser Linux troubleshooting](/tools/browser-linux-troubleshooting)
- [Browser WSL2 troubleshooting](/tools/browser-wsl2-windows-remote-cdp-troubleshooting)