Sources & references
Installation and nightly builds ↗
MiniMind
CodingStudy language model training with a small model, from pretraining and SFT to tool use.
Open sourceLLM training
INDEPENDENT HEADLESS BROWSER
Lightpanda
An independent headless browser for dynamic web extraction and AI automation, with CDP, MCP and script replay.
Official homepage captured in Chrome, not a page rendered by Lightpanda.Explore the image ↓
What you can do with it
What Lightpanda is
Lightpanda is an independent headless browser for web extraction, automation and AI agents. Written mainly in Zig, it executes JavaScript with V8, maintains the DOM and exposes page interactions through the CLI, CDP, WebDriver BiDi and native MCP. It can read dynamic content, extract HTML or Markdown and run fixed workflows.
Version 1.0, released October 2, 2026, ended the Beta phase. The engine does not provide complete graphical page rendering: text PNG/PDF exports come from Markdown layout and cannot replace full CSS screenshots or visual acceptance tests. Validate Web API and client behavior on the target sites.
Local software is open source with AGPL-3.0-only as the default license. Supply your own environment and compute. Cloud meters hosted sessions in browser hours, with model and bring-your-own proxy fees separate. Start with a representative set of authorized pages and measure completeness, success rate, latency and memory before expanding.
01
Dynamic web extraction: Extract lists, details, links and script-generated fields from authorized pages. Define the output schema and sample checks so missing data and error pages are visible.
02
AI web research: Supply Markdown or structured results to a model with source URLs and check dates. Verify model summaries against the original page evidence.
03
MCP agent tooling: Expose navigation, extraction and scripts to an agent, separating tasks through sessions. Handle authorization and outcome checks for login, submission or external writes.
04
Migrate extraction clients: Connect a small Playwright or Puppeteer job first, compare required fields, failures, resources and waiting behavior, then expand page by page.
05
Replay fixed workflows: Export a checked workflow to PandaScript for recurring reading or form prechecks. Keep success conditions, versions and failures, and recheck after page changes.
06
Browser resource evaluation: Compare time, peak memory and completeness on the same sites, workload and outputs. Account separately for local compute, model calls and proxy charges.
A closer look at the product
Official product image. Click to inspect the details.
Project pages and connection examples
Key features, practical tasks
Browser and automation features
- Independent headless engine: Written primarily in Zig, with V8 for JavaScript, Libcurl for HTTP and html5ever for HTML parsing. It is not a Chromium or WebKit fork and focuses on page execution, the DOM and automation.
- Read dynamic pages: Execute scripts and maintain the DOM to read JavaScript-generated text, links and fields. Check each target site for Web API coverage, load timing and authentication; a working homepage does not establish site-wide compatibility.
- HTML and Markdown extraction: The fetch command exports page content. Markdown supplies readable text for models and research, while HTML supports further parsing. Inspect login pages, error pages, pagination and missing fields after extraction.
- Clicks and forms: Use clients or native tools to click, enter text and navigate. Suitable for authorized forms and repeatable workflows. Choose locators from the actual page structure and verify the outcome rather than accepting a completed click as success.
- Playwright / Puppeteer over CDP: Local serve mode exposes CDP over WebSocket. Playwright chromium.connectOverCDP still connects to the independent Lightpanda engine. Protocol support does not imply complete Chrome behavior or API parity.
- WebDriver BiDi connections: WebDriver BiDi is available alongside CDP for compatible Selenium or Puppeteer workflows. Record server protocol options and client versions, then test the commands your task actually uses.
- Native MCP browser tools: Stdio and HTTP servers expose page reading, script execution and session management to MCP clients. Send the Mcp-Session-Id returned at initialization with later HTTP calls; omitting it uses the default shared session.
- Agent planning and PandaScript: The built-in agent can plan browser actions with an LLM and export PandaScript. Fixed scripts can be replayed, avoiding another inference charge for steps without model calls. Site changes and expired sessions still require checks.
- Cookies, requests and proxies: Configure cookies, headers, proxies and request interception for session-based tasks. Proxy providers can charge separately, and a proxy configuration does not grant permission to access or reuse site data.
- Local binaries and Docker: Linux and macOS builds cover x86_64 and arm64; the Docker image is lightpanda/browser. Windows uses Linux through WSL2. No native Windows binary is currently supplied, and Linux binaries depend on glibc.
- Hosted Cloud sessions: Cloud provides remote browser connections with time and concurrency allowances. Browser hours accumulate across open sessions, while concurrency counts sessions open at once. Long tasks need disconnect handling and intermediate results.
- Resource use and measurement: In the published 933-URL, 25-task benchmark, Lightpanda took 4.81 seconds and peaked at 123.0 MB; Chrome took 46.70 seconds and 2.0 GB. The units are processes and tabs respectively. Results do not generalize to every site, and were not rerun here.
- Text PNG / PDF exports: Page Markdown can be laid out as a text PNG or PDF. This is not a screenshot with complete CSS, graphics and font rendering, and is unsuitable for visual regression, pixel comparisons or design acceptance.
- CORS, telemetry and source license: Version 1.0 enforces CORS by default. Telemetry is enabled and can be disabled with LIGHTPANDA_DISABLE_TELEMETRY=true. Source defaults to AGPL-3.0-only; check obligations for deployment, modifications and distribution, separately from Cloud terms.
Get started, step by step
- 01
Choose a build and environment: Install a Linux or macOS build matching the architecture; use WSL2 on Windows. Distinguish stable from nightly, run lightpanda version and record it.
- 02
Read a demonstration page: Start with fetch --obey-robots --dump markdown on the project demo. Inspect text, links and dynamic fields, and check that the result is not an error or login page.
- 03
Start local CDP: Run lightpanda serve --obey-robots --host 127.0.0.1 --port 9222 for local access. The client endpoint is ws://127.0.0.1:9222.
- 04
Connect and validate fields: Create a context and page with playwright-core or puppeteer-core and read one defined field. Add waits, forms and cookies step by step, keeping response status and extracted results.
- 05
Connect an MCP session: Start lightpanda mcp --obey-robots for stdio. HTTP clients retain Mcp-Session-Id after initialization. Use separate sessions for separate tasks and close them afterward.
- 06
Define task success: Check httpStatus, required fields and output types. Handle 404, 500, timeouts, redirects and empty lists; a successful tool call alone does not establish a correct task result.
- 07
Save a replayable workflow: Configure a provider when model planning is needed, export PandaScript and use lightpanda run session.js. Save demonstration data rather than keys or private cookies.
- 08
Measure before scaling: Record success rate, completeness, time and peak memory. For Cloud, track accumulated hours and concurrency allowances too. Increase volume within site permissions and request budgets.
From one-page extraction to a workflow
From one-page extraction to a workflow
Choose a build and environment
Install a Linux or macOS build matching the architecture; use WSL2 on Windows. Distinguish stable from nightly, run lightpanda version and record it.
Read a demonstration page
Start with fetch --obey-robots --dump markdown on the project demo. Inspect text, links and dynamic fields, and check that the result is not an error or login page.
Start local CDP
Run lightpanda serve --obey-robots --host 127.0.0.1 --port 9222 for local access. The client endpoint is ws://127.0.0.1:9222.
Connect and validate fields
Create a context and page with playwright-core or puppeteer-core and read one defined field. Add waits, forms and cookies step by step, keeping response status and extracted results.
Connect an MCP session
Start lightpanda mcp --obey-robots for stdio. HTTP clients retain Mcp-Session-Id after initialization. Use separate sessions for separate tasks and close them afterward.
Define task success
Check httpStatus, required fields and output types. Handle 404, 500, timeouts, redirects and empty lists; a successful tool call alone does not establish a correct task result.
Save a replayable workflow
Configure a provider when model planning is needed, export PandaScript and use lightpanda run session.js. Save demonstration data rather than keys or private cookies.
Measure before scaling
Record success rate, completeness, time and peak memory. For Cloud, track accumulated hours and concurrency allowances too. Increase volume within site permissions and request budgets.
Local use and Cloud costs
LOCAL OPEN SOURCE · OPTIONAL CLOUD
View Cloud pricing ↗Free Local software; Cloud billed separately
Cloud Explorer: ten monthly hours and five sessions free; Builder $19/month. Checked October 15, 2026.
Local open source and Cloud plans
Local open-source use; Cloud has free Explorer and paid Builder / Enterprise plans. Prices checked October 15, 2026.
Explorer — $0/month
Free Cloud allowance; bring-your-own proxy charges are separate.
- 10 browser hours per month
- Up to 5 concurrent sessions
- Community Discord support
Builder — $19/month
Includes 300 browser hours; overage $0.08 per hour.
- 300 browser hours per month
- Up to 30 concurrent sessions
- Email and Slack support
Enterprise — Custom quote
Capacity, support, SLA or private deployment arranged for the workload.
- Custom capacity and service
- Contact the project team
Browser hours accumulate across open sessions. Model, bring-your-own proxy and local compute costs are separate.
A few things worth knowing
Frequently asked questions
Is Lightpanda a desktop browser with an address bar?
It is a headless engine for programs and agents, controlled through the CLI, automation protocols and tool services. Gallery images show the official site and docs in Chrome, not a Lightpanda desktop interface.
Is it still in Beta?
Version 1.0 was released on October 2, 2026, ending the Beta phase. Older articles, forks and nightly builds may describe different states. This profile was checked on October 15, 2026.
Is it a stripped-down Chromium?
It is not a Chromium or WebKit fork. The engine is primarily Zig, with V8 for JavaScript, Libcurl for HTTP and html5ever for HTML. Sharing V8 does not mean sharing the whole Chrome browser implementation.
Does it execute JavaScript?
Yes. It executes JavaScript and maintains the DOM for dynamic content and interactions. Web API coverage still differs, so check actual sites rather than assuming every complex page behaves like Chrome.
Can it replace a browser for visual testing?
Text PNG/PDF exports are laid out from page Markdown, not fully rendered CSS screenshots. Color, font, layout, Canvas and visual regression checks require a browser with the appropriate rendering capability.
Does 9x mean every website is nine times faster?
No. The detailed published test uses 933 URLs, 25 parallel tasks and a specific AWS environment and client. Lightpanda uses multiple processes and Chrome multiple tabs; the single-task gap also differs. We did not rerun it.
Why does the Playwright chromium API connect?
chromium.connectOverCDP is a Playwright CDP client interface. When pointed to the Lightpanda WebSocket, the server engine is Lightpanda. Verify selectors, waiting, cookies and the protocol commands you use.
Is a native Windows exe available?
No native Windows binary is currently supplied. Use the Linux build in WSL2 or a compatible container environment, checking CPU architecture and glibc requirements.
Is every MCP HTTP request an independent session?
No. Send the Mcp-Session-Id obtained during initialization with later requests to continue that separate session. Calls without it use the default shared session. Close sessions when tasks finish.
Does navigation success mean the page is usable?
Do not rely solely on the tool success flag. MCP navigation can return a result for some 404, 500 or timeout cases. Check httpStatus, the target URL and required fields to identify error pages or incomplete content.
Is PandaScript replay entirely free?
A fixed script without LLM calls avoids the corresponding repeated inference charge. It still uses browser compute, networking and possibly paid proxies. Open Cloud sessions consume browser hours.
Does local use require a Cloud subscription?
Local open-source use does not require a Cloud subscription; supply your own environment and resources. Cloud is a separate hosted service with free and paid plans. Check source, Cloud and third-party model terms separately.
How do five sessions and ten hours work on Explorer?
The October 15, 2026 page lists ten monthly browser hours and up to five sessions open at once. Five sessions running one hour consume five hours. Concurrency is not a total API-call allowance; exceeding it returns 429.
Can a Cloud task stay connected indefinitely?
Current documentation lists a 15-minute CDP connection limit and a five-minute MCP idle limit. Save intermediate results, close unused sessions and handle disconnects. Free usage stops at the monthly allowance; paid overage is metered.
How should local data and telemetry be handled?
Telemetry is enabled by default and can be disabled with LIGHTPANDA_DISABLE_TELEMETRY=true. Model providers, search and external pages involve separate network communication. Private pages, logs and scripts can contain cookies or content; review configuration and retention.
Which source license applies?
The LICENSING file specifies AGPL-3.0-only by default. Check source and notice obligations for modifications, distribution and network services. Cloud and third-party components have separate terms; open source does not mean unconditional use.
The details behind your decision
Explore platforms, inputs and outputs, licensing, and access requirements.
Environment and entry commands
Environment and entry commands
These commands represent separate entry points, not an installation script to execute line by line. Use ./ for a binary in the current directory; use lightpanda after adding it to PATH.
./lightpanda version
./lightpanda fetch --obey-robots --dump markdown https://demo-browser.lightpanda.io/campfire-commerce/
./lightpanda serve --obey-robots --host 127.0.0.1 --port 9222
chromium.connectOverCDP('ws://127.0.0.1:9222')
lightpanda mcp --obey-robots
lightpanda mcp --host 127.0.0.1 --port 9223 --obey-robots
lightpanda agent --no-llm
lightpanda run session.js
Sessions, data and local services
Sessions, data and local services
Telemetry is enabled by default and can be disabled with LIGHTPANDA_DISABLE_TELEMETRY=true. Model providers, search and external pages involve separate network communication. Private pages, logs and scripts can contain cookies or content; review configuration and retention.
Keep default CORS checks and bind workstation services to 127.0.0.1. Run tasks within site permission and request budgets; --obey-robots enables robots checks.
Sources, documentation and help
Project resources and further reading
Project resources and further reading
Checked October 15, 2026 against public source, the 1.0 announcement and official documentation. The browser and benchmark were not run for this profile.
- Project repository
- Official website
- Version 1.0 announcement
- Installation and nightly builds
- Playwright CDP
- MCP and sessions
- Agent mode
- PandaScript
- Browser concepts
- Benchmark method
- Cloud pricing
- Cloud limits and billing
- Licensing
- Privacy policy
- Our Lightpanda project introduction
- Lightpanda setup guide
Your experience can help someone choose
COMMUNITY
Sign in to rate and comment
Tried it? Share your experience
Loading experiences…
Manage my comments
Community and moderation
Saved tools are private. Approved comments are public; edits return to moderation. Comment counts include approved comments and replies. Each person has one rating. New ratings are automatically approved; withdrawn, unapproved or invalid ratings do not count. Share real experiences and avoid spam, private information or personal attacks.
For moderation appeals or data requests: support@findgoodai.com
Keep exploring similar tools
OpenRouter
CodingCall multiple AI models through one API and choose provider, fallback and privacy routing rules. OpenRouter passes through model prices, charges fees when purchasing credits, and applies a plan-dependent allowance then a 5% BYOK fee.
Unified model APIProvider preferences
Gemini CLI
CodingUse Gemini CLI to work with code and files from a terminal through a supported API key or enterprise account. Google moved individual free and Pro/Ultra account access to Antigravity CLI in June 2026; check the authentication route before installing.
Terminal code explorationCode and test edits
RELATED ARTICLES
Related articles
Getting started with Lightpanda: local setup, Playwright and MCP sessions
A practical path from version selection and Markdown extraction to CDP and MCP connections, with response checks, Cloud billing and text-image export limits.
NEXT STEP
Explore Lightpanda ↗Start with one extraction result you can verify
Bring a real task and see how it fits the way you work.
Jump to a section
User experiences