Unbrowse: the drop-in browser for OpenClaw agents on websites

Unbrowse is an open-source CLI that turns websites into reusable skills for agents by learning the request path behind the page and replaying it later.

Setup: curl -fsSL https://unbrowse.ai/install.sh | bash

Optional skill install: npx skills add unbrowse-ai/unbrowse

MCP setup: git clone --single-branch --depth 1 https://github.com/unbrowse-ai/unbrowse.git ~/unbrowse && cd ~/unbrowse && ./setup --host mcp

Upgrade existing installs: curl -fsSL https://unbrowse.ai/install.sh | bash

MCP config template: https://www.unbrowse.ai/mcp.json

Docs: https://docs.unbrowse.ai

Usage: unbrowse resolve --intent "..." --url "..." --pretty

Free, open source, local-first

For OpenClaw agents on real websites

Make your OpenClaw agent 100x faster,90% cheaper on websites.

Install one plugin. Skip screenshots, DOM waits, and selector repair. Unbrowse uses the browser once to learn the useful path, then reuses it on the next run.

Built for agents that currently depend on Playwright-style browser loops to reach real websites. Good routes compound into reusable skills, and shared reuse can turn mined routes into contributor income.

Also works with Claude Code, Codex, Cursor, and MCP hosts.

94 domains benchmarked
3.6x faster than Playwright on average
5K+ npm downloads

Optional Host Shortcut

After install. Adds slash-command or host discovery in skill-aware tools.

npx skills add unbrowse-ai/unbrowse
Drop-in OpenClaw browserFull install matrixSee live flowDocs

Browser tax vs route reuse

Same website. Less browser.

replay-ready

Browser automation

Re-do the website

Open page1.8s
Wait for DOM4.9s
Click + scrape9.3s
Parse output14.6s

Unbrowse skill

Replay the useful path

Resolve skill42ms
Call route118ms
Return data184ms

What ships as the skill

Not just a URL. Auth path, parameters, schema, and enough reliability state to keep the task useful.

unbrowse resolve --intent "get trending searches" --url "https://google.com" --pretty
auth-awarestructured outputrefreshablebrowser fallback

50-200ms

direct replay

~200

tokens / action

local

auth + cookies

100x faster

Skip screenshots and DOM waits.

OpenClaw usually pays the full browser tax on every run. Unbrowse learns the useful path once and reuses it on the next task.

90% cheaper

Pay for the task, not the browser.

Return structured data or actions without shipping a full browser loop through every run.

Compounds

Mine routes that can earn passive income.

Mine the internet into reusable skills. If other agents keep reusing what you mined, contributor payouts can compound from that reuse.

Why OpenClaw Users Switch

Replace the browser loop. Keep the website task.

Unbrowse is for agents that need to log in, click through real sites, and bring back data or actions. It is not for generic browser QA, pixel tests, or full end-to-end UI suites.

Old way: Playwright, screenshots, waits, selectors, retries.

New way: install Unbrowse as the OpenClaw browser layer and reuse the learned path.

When the site shifts: re-browse once, refresh the route, keep going.

What compounds

The first good run pays for the next ones.

One working route removes a lot of repeated browser work. When you publish a route other agents reuse, that shared reuse can turn into contributor income.

Keeps login working

A useful skill keeps the authenticated path intact while still using the browser for login, refresh, and fallback when needed.

Returns data, not markup

The output is a callable interface with inputs and structured responses, not another HTML page your agent has to parse.

Refreshes when sites move

When a route degrades, Unbrowse can re-browse, refresh the execution plan, and keep the task stable.

Compounds on reuse

Once a route is good, it stops being one-off automation and becomes infrastructure another agent can call again.

Why teams switch

5-30s

browser replay

50-200ms

skill replay

~200

tokens / action

Not for generic browser QA. Very much for OpenClaw and agent workflows on real websites, especially when mined routes can be reused by others.

Replace The Browser

Install the plugin. Then watch one real website task work.

Fastest way to pattern-match the product: install Unbrowse into OpenClaw, run one website task, then watch the next run skip most of the browser loop.

Default path

curl -fsSL https://unbrowse.ai/install.sh | bash

Start with the runtime. Add the skill or MCP wiring only after the local path is healthy. If you want the contributor upside, keep the mined routes you publish useful enough that other agents actually reuse them.

CLICodexClaude CodeCursorMCPOpenClaw

Canonical quickstart

unbrowse resolve --intent "get trending searches" --url "https://google.com" --pretty
1
2
3
4
5
6
7
8
9
10
11
12
13
# One-line install
curl -fsSL https://unbrowse.ai/install.sh | bash
# During first run, set up Crossmint lobster.cash if you want route-mining payouts
# Optional: add the skill after install
npx skills add unbrowse-ai/unbrowse
# Verify
unbrowse health --pretty
# Upgrade after releases
curl -fsSL https://unbrowse.ai/install.sh | bash
Demo

One agent mines Airbnb once. The next agent just uses the skill.

This is the product in one motion: mine the route, package it, replay it.

unbrowse
Agent
Online
Unbrowse airbnb.com — I need to search listings
2:41 PM
Opening headless browser, navigating airbnb.com...Captured 12 API endpoints across 4 services
GET/api/v3/StaysSearch- Search listings
GET/api/v3/StayListing- Listing details
GET/api/v3/PdpAvailabilityCalendar- Availability
GET/api/v3/StaysPriceBreakdown- Price breakdown
+ 8 more endpoints
Skill generated: airbnb-stays-api (quality: 91/100)Automatically indexed — every agent on the network can now use this
2:42 PM
Find me places to stay in Tokyo for 2 guests, March 15-22
2:43 PM
Replaying airbnb-stays-api → GET /api/v3/StaysSearch
Executed200 OK · 0.4s
listings found:243 results
top result:Shibuya Loft with City View
price:$89/night
rating:4.92 ★ (318 reviews)
243 listings found — $89-$420/night range
2:43 PM
Can a different agent use this too?
2:44 PM
Yes — the Airbnb skill is already in the shared index. Any agent can search for it by intent and replay the endpoints directly. No browser needed, no re-discovery.
One agent browses. Every agent knows.
2:44 PM
Message your agent...
FAQ

Questions skeptical builders actually ask

Who is Unbrowse for?

It is for OpenClaw users and agent builders whose agents do repeated tasks on real websites and are tired of paying the full browser cost every time.

How does Unbrowse work?

Use the browser once to learn the website task. After that, Unbrowse can replay the learned path as a reusable skill instead of forcing the agent to click through the whole UI again.

How much faster is it than headless browser automation?

Across 94 benchmarked domains, the route-reuse path was 3.6x faster on average than Playwright. On repeated workflows, the gap is often bigger because Unbrowse avoids the full browser replay loop.

Is this just scraping?

No. Scraping gives you page output. Unbrowse is trying to recover a callable execution path - something an agent can run again with auth, parameters, and structured responses.

Do credentials leave my machine?

No. Unbrowse runs locally, keeps browser-backed auth local, and does not depend on a cloud proxy or man-in-the-middle layer.

What if the website changes its internal API?

When a learned route goes stale, Unbrowse can re-browse the site, capture the updated flow, and refresh the skill instead of forcing you to rewrite selector-heavy scripts.

How do I install it?

Start with curl -fsSL https://unbrowse.ai/install.sh | bash. That script installs the CLI, runs setup, and finishes the real first-run flow: ToS acceptance, agent registration plus API-key caching, and wallet detection when present. After install, hosts with skills support can also use npx skills add unbrowse-ai/unbrowse for slash-command or host discovery. For generic MCP hosts, run git clone --single-branch --depth 1 https://github.com/unbrowse-ai/unbrowse.git ~/unbrowse && cd ~/unbrowse && ./setup --host mcp; that writes a ready-to-import config to ~/.config/unbrowse/mcp/unbrowse.json, and the generic template lives at https://www.unbrowse.ai/mcp.json. Upgrade CLI installs with curl -fsSL https://unbrowse.ai/install.sh | bash and MCP installs with cd ~/unbrowse && git pull --ff-only && ./setup --host mcp. For OpenClaw, install the published browser-replacement plugin with npx unbrowse-openclaw install --restart.

Does it work with OpenClaw?

Yes. Unbrowse can act like a drop-in browser replacement for OpenClaw, so the agent reaches websites through reusable skills instead of the normal browser loop when that path is available.

What actually gets shared?

The point is to share the route knowledge and execution plan, not your personal session data. The reusable asset is the skill.

Can I earn passive income from mined routes?

That is the contributor upside. If you mine useful routes and other agents keep reusing them, contributor payouts can accrue from that reuse instead of the work staying trapped in one local session.

What is it not for?

If your main job is browser QA, visual regression testing, or full end-to-end UI testing, Playwright is still the better fit. Unbrowse is for agent execution on websites, not generic browser testing.