unbrowse

name	unbrowse
description	Web content + endpoint discovery without the browser-driving overhead. Use for clean markdown fetch of any URL (libcurl-impersonate Chrome JA4, auto-pulls cookies from your real browser), one-shot intent-driven scraping (`unbrowse run`), and capture/replay of authenticated JSON endpoints so you don't have to drive a UI every time. Primary surface is the CLI (`unbrowse <cmd>`); MCP server is available at `npx -y unbrowse mcp` as an alternative for hosts that have it wired up.
user-invocable	true
metadata	{"openclaw":{"requires":{"bins":["unbrowse"]},"install":[{"id":"npm","kind":"node","package":"unbrowse","bins":["unbrowse"]}],"emoji":"🔍","homepage":"https://github.com/unbrowse-ai/unbrowse"},"version":"7.1.0","updated":"2026-05-27"}

Unbrowse — Web fetch, intent-driven scrape, endpoint capture/replay

Unbrowse is a CLI (and optional MCP server) that gives an agent three things the regular fetch tool can't:

1. Realistic HTTP fingerprint — every request goes through libcurl-impersonate (Chrome 131 JA4) and auto-pulls cookies from your actual installed Chrome/Edge/Brave/etc. profile. Most "WAF-blocked when scripted" pages just work.
2. HTML → clean markdown by default — unbrowse fetch <url> returns body-only readable text, no nav chrome, no scripts. Better signal-to-noise than raw curl | pandoc.
3. Capture-and-replay of authenticated JSON endpoints — unbrowse run (or resolve + execute) opens a real browser exactly once, discovers the API call behind a page, caches it, and replays it as a plain HTTP request on every subsequent run. Skills accumulate in a local index keyed by domain + intent.

When to use unbrowse

Use unbrowse for	Use something else for
Clean-markdown read of any public URL → unbrowse fetch <url>	Reading a local file → Read tool
Page that's WAF/cookie-gated and curl returns 403	Page in your already-open Chrome tab → chrome-devtools MCP at :9222
First-time scrape of an authenticated UI ("download my X from Y") → unbrowse run <url> "task"	Sending email → gog send -a <addr>
Discovering the JSON API behind a page → unbrowse resolve --intent "..." --url "..."	Bulk Composio integrations → Composio CLI
Replaying a captured endpoint with new params → unbrowse execute --skill ID --endpoint ID -p key=val	Greping the codebase → Grep/Glob
Live HAR capture while you click around → unbrowse capture --url ... --intent ...	Image/PDF inspection → Read tool

When NOT to use unbrowse

• Don't use to extract cookies for another tool. Use ~/tools/cookies-txt or chrome-devtools CDP per the cookie-handling rules.
• Don't use as a Grep substitute. It fetches the network — Grep/osgrep is for local files.
• Don't use for image or PDF content. The Read tool handles those natively.
• Don't use for already-open browser tabs. chrome-devtools MCP at :9222 is faster and sees the live DOM state.

The 5 commands an agent uses 95% of the time

1. unbrowse fetch <url> — clean markdown of any URL

unbrowse fetch "https://example.com/article"
# → body-only markdown, no nav, no scripts, cookies from your real browser

Flags worth knowing: --raw keeps HTML bytes (no markdown conversion). For sandboxed JS evaluation in the page context, use the advanced mode: unbrowse fetch <url> --bundle-source <js|-> --post-eval <expr> which returns the full envelope (cookies, post_eval result, observed routes).

2. unbrowse run <url> "<task>" — one-shot intent

unbrowse run "https://sanfrancisco.nextrequest.com" "list my open records requests"

First call: opens a real browser, captures the API. Subsequent calls: replays cached endpoint. Falls back to interactive only when needed.

3. unbrowse resolve --intent "..." --url "..." [--no-execute] — endpoint discovery

unbrowse resolve --intent "fetch latest releases" --url "https://github.com/owner/repo"
# → auto-executes the top-ranked safe GET. Add --no-execute for just the shortlist.

Use --no-execute to inspect what's available, then unbrowse execute --skill ID --endpoint ID -p key=val to actually call it.

4. unbrowse capture --url <url> --intent <intent> — explicit live HAR

unbrowse capture --url "https://app.example.com/dashboard" --intent "download monthly report"

Opens a visible browser, you do whatever you'd do manually, unbrowse logs the requests and indexes them as replayable endpoints.

5. unbrowse auth <url> — first-time login

unbrowse auth "https://app.example.com"
# Opens a visible browser. You sign in. Cookies persist for all future fetch/run/resolve.

Quick global flag reference

Flag	What it does
--pretty	Pretty-print JSON output (indented)
--raw	On fetch: keep HTML bytes (skip markdown). On resolve/execute: skip server-side projection
--no-auto-start	Don't auto-spawn the local server if it's down
--path "data.items[]"	Drill into the result before extract/output (resolve/execute)
--extract "field1,alias:deep.path"	Pick specific fields without piping
--limit N	Cap array output to N items
--dry-run	Preview mutations without applying

Full command/flag dump (auto-generated, kept here for completeness):

All commands

Command	Description
fetch <url>	Primary URL → markdown. Add --bundle-source <js> + --post-eval <expr> for sandboxed JS evaluation.
run <url> "task"	One-shot intent. Cached replay first, capture+index on miss, browser only when needed.
resolve --intent "..." [--url "..."]	Resolve intent against marketplace + local cache. Auto-executes top safe GET; --no-execute for metadata only.
execute --skill ID --endpoint ID -p k=v	Call a specific endpoint after resolve --no-execute.
capture --url <url> --intent "..."	Live-browser HAR capture; discovers + indexes API endpoints.
auth <url>	Open a visible browser for sign-in; cookies persist for future calls.
explain --intent "..." --url "..."	Print top-N candidate endpoints + evidence (no verdict, just primitives).
note <read|write|list> --domain <d>	Per-domain LLM notes consumed by augment on next capture.
skills	List all locally-cached skills.
skill <id>	Full SkillManifest for one skill.
feedback --skill ID --endpoint ID --rating 1-5	Submit feedback after presenting results (mandatory after resolve+execute).
annotate --skill ID --endpoint ID --text 'tip'	Contribute best practices/gotchas.
review --skill ID --endpoints '[...]'	Push reviewed descriptions/schema back to a captured skill.
index --skill ID	Recompute local graph/contracts (no network).
publish --skill ID	Publish reviewed skill to marketplace (gated by unbrowse mode).
cleanup-stale [--skill ID]	Verify skills against live endpoints; evict stale entries.
health	Local server health check (version, uptime).
account [--register] [--json]	Show local account, wallet, contribution mode.
mode	Re-prompt for private / share / share+earn.
settings	Show/update local capture/publish policy.
dashboard	Open website dashboard, pair this CLI install.
setup	Bootstrap browser engine + /unbrowse command. Idempotent.
upgrade	Print the right upgrade command.
stats [--flywheel | --earnings]	Time/tokens/cost saved + earnings.

Browse-session commands (when run falls back to interaction)

go <url>, snap, click <ref>, fill <ref> <value>, type <text>, press <key>, select <ref> <value>, scroll <dir>, submit, screenshot, text, markdown, cookies, eval <expr>, back, forward, sync, close, inspect, sessions --domain <d>.

resolve/execute extras

Flag	Description
--no-execute	Resolve only; return shortlist without auto-executing
--schema	Show response schema + extraction hints (no data)
--endpoint ID	Pick a specific endpoint by ID
--params '{...}'	Extra params as JSON
-p key=val	Single param via repeated flag
--require-proof	Filter to endpoints with independently verified proofs

MCP option (alternative to CLI)

If your host has it wired up, npx -y unbrowse mcp exposes the same capabilities as MCP tools (unbrowse_resolve, unbrowse_execute, unbrowse_fetch, unbrowse_run, unbrowse_capture, plus the browse-session tools unbrowse_go/snap/click/etc., and unbrowse_publish/unbrowse_review for the marketplace flow). Result size is capped at 25 KB per call.

Add to your MCP host config (Claude Desktop, Cursor, Codex, etc.):

{ "mcpServers": { "unbrowse": { "command": "npx", "args": ["-y", "unbrowse", "mcp"] } } }

The CLI is the same binary that backs MCP, so any host with shell access already has full unbrowse — MCP just makes the surface discoverable as typed tools.

Help

• Repo: https://github.com/unbrowse-ai/unbrowse
• Docs: https://unbrowse.ai
• Issues: https://github.com/unbrowse-ai/unbrowse/issues

---

NOTE — historical: Earlier versions of this skill (≤ v6.14.0) declared themselves deprecated in favor of MCP-only. That guidance is reversed as of v7.1.0 (2026-05-27): the CLI is the primary agent surface, the MCP server is an equal alternative, and shell-only environments (subagents, CI, /carmack, /debug) should reach for unbrowse <cmd> directly.