Lightweight browsers
Lightpanda is an opt-in engine for text and DOM browser tasks. It uses the same OpenClawbrowser tool through an explicitly configured profile. It is not a
visual-browser replacement: keep a Chromium profile for screenshots, PDF output,
and applications that require unsupported browser features.
The examples pin Lightpanda 0.4.1. They do not change your existing browser
profile, install a service, or migrate a logged-in Chrome profile.
Browser plugin architecture
The bundled Browser plugin owns both engine adapters. Chromium and Lightpanda use the samebrowser tool, profile selection, route admission, navigation
policy, and session lifecycle. The registered adapter selects capabilities,
CDP normalization, snapshot defaults, and connection lifetime; it does not add
a second browser tool or process manager.
Browser status reports availableEngines and the selected engine’s
sessionScope and screenshotFidelity. Chromium keeps its managed,
existing-session, extension-relay, and remote-CDP profiles. Lightpanda remains
an explicitly selected, attach-only external engine with one page per
connection and no automatic read replay after session loss.
These are adapters inside the existing plugin, not separately installable
third-party plugins. Enabling the adapter does not download, launch, or bundle
the Lightpanda engine. Existing profiles and engine configuration are unchanged.
Licensing and distribution
OpenClaw’s adapter remains MIT-licensed. The optional Lightpanda engine is AGPL-3.0-or-later, not MIT: see its pinned source notice and license. The existing Playwright client is Apache-2.0; the existingws client is MIT.
Their licenses and third-party notices still apply.
These examples connect over CDP to an independently installed, unmodified engine.
They do not bundle or relicense Lightpanda in OpenClaw’s package or image.
The engine and its container dependencies are not an MIT-only distribution.
If your deployment excludes copyleft software, do not select this engine.
Mirroring or bundling the engine requires a separate redistribution review,
including license notices, Corresponding Source, and third-party obligations.
Modifying a network-served engine also requires reviewing AGPL section 13.
A separate process is not a blanket legal exemption, and checksum verification
does not establish license compliance. See the
artifact and dependency review
for the verified pins and remaining limits.
Alternatives reviewed
The following is a licensing comparison as of 2026-09-21, not a claim that these alternatives have passed OpenClaw integration or cross-platform tests. An MIT-compatible application and an entirely permissive engine distribution are different requirements.
Automation clients do not replace the engine. For example,
Vercel’s agent-browser
offers Chrome and Lightpanda backends; its own Apache-2.0 license does not change
the selected engine’s license. The same distinction applies to Playwright and
Puppeteer clients.
Obscura audit boundary
Release v0.2.3 was reviewed at commit1a3169da276d7720732c7b20535474942917fb83.
Its lockfile
contains 471 registry packages. Their license declarations include no AGPL,
but five are MPL-2.0-only: cooked-waker 5.0.0, cssparser 0.34.0,
cssparser-macros 0.6.1, dtoa-short 0.3.5, and selectors 0.26.0.
These dependencies remain in the no-render engine’s DOM/JavaScript paths.
MPL’s file-level obligations
do not require unrelated MIT adapter files to become MPL; they still apply to
the covered code and its distribution.
The downloaded Linux x86-64 no-render archive matched release SHA-256
b5e55e8f2c97814127a521cd59af1a84b79dc40cf658fda04df04af81a2d89f3.
It contained only obscura and obscura-worker, without license or notice files.
The pinned release workflow
packages only those executables and does not use Cargo’s --locked flag.
Consequently, source-lock metadata is not proof of the complete dependency set
inside each release binary. Full V8/third-party and platform-library review,
notices, source availability, and runtime compatibility remain to be verified
before bundling or recommending an integrated deployment.
Choose where the engine runs
localhost inside an OpenClaw container means that container, not the host and
not the Lightpanda sidecar. Use the service name for container-to-container
connections. On Windows, run Docker Desktop in Linux container mode, or run
both OpenClaw and the Linux engine inside WSL. Lightpanda does not publish a native
Windows binary. macOS and Linux have official x86-64 and ARM64 release binaries;
the official container image has Linux amd64 and arm64 variants.
Upstream installation information.
Docker with OpenClaw on the host
From the repository root:127.0.0.1:9222 only. Set LIGHTPANDA_PORT to select
another host port, and update the profile URL to match. The image is pinned by its
multi-platform digest, so Docker selects the host architecture without pulling a
moving latest or nightly version.
CDP gives a client control over the browser; the sample does not add CDP
authentication. Do not change the loopback binding to a public address. Use an
authenticated tunnel for access from another host.
To stop and remove only this sample’s container and network:
Docker Compose with OpenClaw in a container
Merge the sidecar into the repository’s existing Compose project:cdpUrl: "ws://lightpanda:9222" in the profile
below. Use the same Compose files and project name when starting the Gateway.
No browser port is published to the host in this variant. The containers share
the project’s bridge network and retain outbound internet access; the network
is not declared internal: true because that would prevent public-site browsing.
Use your normal OpenClaw Docker setup for its state directory, authentication,
and Gateway startup. The sidecar does not mount your OpenClaw state, browser
cookies, or host Docker socket.
For Podman, use an installed Compose provider and verify service-name DNS before
choosing the sidecar URL. A netavark installation without its aardvark-dns
helper can start a loopback-published engine while leaving container DNS broken;
successful engine startup does not prove sidecar connectivity.
Native Linux and macOS
Download the release binary for your operating system and CPU from Lightpanda 0.4.1. The sample’sdeploy/lightpanda/SHA256SUMS records the release asset digests.
For Linux x86-64, run from the repository root:
lightpanda-aarch64-linux for Linux ARM64 or
lightpanda-x86_64-macos for Intel macOS. Only execute the downloaded binary after
its checksum matches. Linux release binaries require glibc; use the official
container image on musl-based systems such as Alpine. These commands run the
engine in the foreground; stop it with Ctrl+C.
Configure an opt-in profile
Merge this browser block into your existing configuration:profile: "lightpanda" on browser tool calls. When the selected workload has
passed your checks, set browser.defaultProfile to "lightpanda" to make it the
default. Preserve your Chromium profile and select it explicitly for visual or
unsupported work. Restore the previous defaultProfile to undo the selection.
engine declares the capability contract; a CDP endpoint alone does not imply
Chromium compatibility. attachOnly means OpenClaw attaches to the service you
started instead of launching or taking ownership of a local Chrome process.
Do not set executablePath to Lightpanda: its CLI is not Chrome’s launch CLI.
Session and capability limits
- A Lightpanda CDP connection owns its page state. Closing the connection, stopping the container, or restarting the engine loses that state; reconnecting does not resume the previous page or login.
- One CDP connection supports one page target. Separate connections can coexist, but a Lightpanda profile is not a general multi-tab Chromium session.
- No automatic cross-engine replay occurs after an action fails. A click or form submission may already have happened; inspect its outcome before repeating it.
- Lightpanda’s text-layout preview is not a rendered screenshot. It cannot prove CSS, image, font, or visual-layout correctness.
- The verified snapshot path is AI format with
ariareferences. The engine selects those references by default, including efficient snapshot mode. Explicit role references, selector/frame-scoped snapshots, labeled screenshots, and the separateariasnapshot format are unsupported in this adapter. - JavaScript and web APIs are not a guarantee that every website will work. Verify the sites and interaction patterns you actually use.
Verification and benchmarks
Engine startup, CDP connectivity, task completion, and full OpenClaw integration are separate checks. A running container or a successfulBrowser.getVersion
does not prove that snapshots, references, and actions work through OpenClaw.
Chromium headless shell baseline
For an alternative without Lightpanda’s AGPL engine, first test Chromium’s headless shell through the existing Chromium profile. It retains Chromium’s third-party license obligations; this is not an MIT-only binary. It does not require another automation daemon or an OpenClaw engine adapter. Use the repository-pinned Playwright installer rather than an unpinned wrapper:chrome-headless-shell (or chrome-headless-shell.exe on
Windows) in that directory. Linux also needs the browser’s system libraries and
fonts; see Linux troubleshooting.
Run from the repository root, quoting paths that contain spaces:
--headless-shell selects
the benchmark executable only: it does not install a production browser,
change a profile, or establish binary provenance. Preserve its complete
distribution and LICENSE.headless_shell when reviewing deployment. The installer
also downloads platform helper assets, including FFmpeg; review and retain their
own notices separately. Use
separate invocations for the full Chromium and headless-shell comparisons;
memory or startup savings must be measured, not inferred from download size.
Native engine comparison
Run the opt-in synthetic route benchmark from the repository root after installing development dependencies:LD_LIBRARY_PATH and FONTCONFIG_FILE without changing the host’s packages.
Lightpanda additionally checks unsupported-operation rejection, its single-page
limit, and stale-target rejection after disconnecting. These checks do not use
an LLM and do not measure model reasoning or end-to-end agent token cost.
--iterations accepts 1 through 100 and counts warm tasks. A separate first
task includes the initial page open and CDP attachment; every warm task includes
navigation and the same form workflow. Native runs also report process startup
and time from startup through the first completed task. Warm percentiles exclude
the first task. Capability/session checks run after the measurement window.
A combined run uses one Node controller and records engine order; its later
engine can reuse controller modules already loaded by the earlier engine. The
first-task and process-start figures are not cold CLI/controller measurements.
Use separate invocations when comparing independently initialized controllers.
The memory fields are maximum sampled process-tree PSS/RSS, not true peaks.
They use Linux /proc with sampling attempts every 50 ms and at task boundaries.
Short-lived processes or transient allocations can be missed. Unsupported hosts
and externally managed engines report null, as do runs with unreadable process
memory, never a guessed engine-memory
figure. Controller RSS is a separate end-of-workload sample, not incremental
controller overhead; do not add independently sampled maxima and call the sum
total peak host memory.
An externally managed engine
Use a dedicated engine instance. External mode closes the benchmark’s control connection and its own Chromium tab, but does not stop the engine process:--fixture-host must name the controller from the browser’s network, not
from the controller’s own network. Linux Docker needs a reachable host address
or a configured host-gateway mapping. The default fixture listener/hostname
remain 127.0.0.1 when those flags are omitted.
External mode reports engine startup and memory as null. Capture the pinned
container/binary version separately with the report. A Windows Node controller
can use the same external-engine interface, but Windows Docker runtime behavior
has not been verified for this sample.
Compare the same deterministic tasks with a pinned Chromium baseline. Record
task completion before reporting speed or memory improvements; unsupported or
failed work must not be counted as a successful fast result. Report warm and
cold runs separately, engine versions, host OS/architecture, client overhead,
and whether memory includes the whole process tree or container.
For Docker Desktop, container memory does not include the VM’s host overhead.
Do not compare a native-process RSS figure with a container-only figure and call
the difference total host savings. Keep all benchmark fixtures public or local;
do not export an existing logged-in browser profile to make a benchmark pass.
Platform support listed above describes upstream distribution and the deployment
topologies, not a claim that every platform has passed the same runtime tests.
Record actual platform and container-runtime results with the benchmark report.