unbrowser
Agent-native browser and MCP server for lightweight, Chrome-free web discovery, with explicit escalation to Unchained for browser-only flows.
Documentation
unbrowser
The cheap browser pass for agents. One native binary. No Chrome.
Official MCP Registry identity: `mcp-name: io.github.protostatis/unbrowser`
`unbrowser` is a stateful, non-visual web runtime for agents. It sits between `curl`/WebFetch and a real browser: it retains cookies and DOM state, returns queryable element refs, handles links and HTML forms, detects challenge and SPA signals, and can run bounded QuickJS page scripts when static HTML is not enough.
> Default `navigate` is a fast static/SSR pass. Set `exec_scripts: true` only when you need bounded QuickJS execution. Heavy SPAs, pixels, V8 fidelity, extensions, and interactive anti-bot challenges belong in a real Chrome tier.
Try the live public-web demo before installing. It accepts only the fixed public source sets shown on the page; do not send private data, cookies, or authenticated tasks through it. A shared Streamable HTTP MCP endpoint is available at `https://unchainedsky.com/unbrowser-mcp` for public smoke tests; production sessions should use a local install.
Use `unbrowser` only on sites, accounts, and data you are permitted to access.
Challenge detection and cookie import do not grant permission to bypass access
controls or site terms. Treat session cookies as credentials: keep them local,
scope them to the authorized task, and never send them to the public demo.
Pick the right tier
| Need | curl / WebFetch | unbrowser | Real Chrome |
|---|---|---|---|
| Static / SSR HTML | raw response | structured BlockMap + DOM queries | full browser |
| Cookies, links, HTML forms | DIY | built in | built in |
| Client-side page scripts | ❌ | bounded QuickJS, opt-in | V8 |
| Pixels, Canvas, WebGL, Workers, extensions | ❌ | ❌ | ✅ |
| Agent-oriented output | DIY parsing | element refs, page signals, structured extraction | DIY CDP / DOM parsing |
| Interactive access challenge | ❌ | detect + stop or escalate | authorized browser / human confirmation |
Use `unbrowser` when HTTP alone is too dumb and a full browser is too expensive. When the page needs Chrome, the output tells the agent to escalate rather than pretending compatibility it does not have.
Quick start
Docker — Linux amd64/arm64, ~13 MiB pull
docker pull ghcr.io/protostatis/unbrowser:latest
# One-shot navigation
docker run --rm ghcr.io/protostatis/unbrowser:latest \
navigate https://example.com --json
# Default mode: MCP over stdio
docker run --rm -i ghcr.io/protostatis/unbrowser:latestThe image is distroless and runs as non-root: no shell, package manager, or persistent state. Pin `:vX.Y.Z` or an image digest in production.
Python
pipx install pyunbrowser # recommended on macOS / modern Linux
# or, inside a Python 3.10+ virtual environment:
pip install pyunbrowserfrom unbrowser import Client
with Client() as ub:
ub.navigate("https://news.ycombinator.com")
for link in ub.query(".titleline > a")[:3]:
print(link["text"], link["attrs"]["href"])On macOS, `/usr/bin/python3` is 3.9 and cannot install the wheel; use `pipx` or a Homebrew Python. The PyPI distribution is `pyunbrowser`, while the import and executable remain `unbrowser`.
MCP
{
"mcpServers": {
"unbrowser": {
"command": "unbrowser",
"args": ["--mcp"]
}
}
}Minimal surface: add `"--mcp-profile", "minimal"` to expose only `navigate`/`query`/`extract`/`help` and let the agent discover the rest via `help(topic)`. The Python wrapper ships a 3-tool smart server (`search`/`open`/`help`) as the `unbrowser-smart` console script.
Docker MCP configuration
{
"mcpServers": {
"unbrowser": {
"command": "docker",
"args": [
"run", "--rm", "-i",
"ghcr.io/protostatis/unbrowser:latest"
]
}
}
}See installation and interface reference for Cargo, release archives, source builds, persistent shell sessions, raw JSON-RPC, and all MCP options.
What an agent gets
- `navigate` returns a BlockMap: page title, landmarks, headings, interactives, density signals, and an ASCII outline. Its size is page-dependent; it is structured for planning rather than a fixed-token promise.
- Stable element refs (`e:142`): query an element once, then `click`, `type`, or `submit` it without re-parsing HTML.
- Stateful cookies and forms: cookie jar, GET and URL-encoded POST form submission, links, and redirects persist within a session.
- Page and challenge signals: `density.likely_js_filled`, `thin_shell`, and `challenge.provider` tell an agent whether to run scripts, inspect embedded data, stop, or escalate to an authorized browser session.
- Routing aids on every result: `micro_hint` (the single next concrete step), `next_tools` (ranked candidates), `avoid` (tools with nothing to act on), and a stable `escalation` taxonomy (`challenge`, `thin_shell`, `partial_result`, …) with retryable/severity/evidence — recommendations never contradict `avoid`.
- Structured helpers: route discovery, card extraction, table normalization, `text_main`, and selector debugging cover common extraction workflows.
Script mode and escalation
{"id":1,"method":"navigate","params":{"url":"https://example.com","exec_scripts":true}}With `exec_scripts: true`, inline and external scripts run in QuickJS under a bounded watchdog. This can materialize light hydration and fetch-visible data; it is not V8 or a rendering engine. Heavy React/Vue/Ember apps may still leave an empty shell.
Escalate to `unchainedsky-cli` or Unchained when a permitted task needs real pixels, Canvas/WebGL, Workers, browser extensions, V8 compatibility, an authenticated profile, or human confirmation. When the user has already established an authorized session in Chrome, `cookies_set` can import the required session cookie locally until it expires; it is not a license to circumvent a site's controls.
Documentation
| Need | Read |
|---|---|
| Install paths, session CLI, one-shot CLI, raw RPC, MCP, shims, full RPC table | Usage reference |
| Script compatibility, SPA signals, challenge handling, authorized cookie handoff, escalation | Compatibility and escalation |
| Distribution and supported directory listings | Distribution notes |
| Build the native binary | Build instructions |
License
Apache-2.0 — see LICENSE.
Frequently asked questions
What is unbrowser?
unbrowser is Agent-native browser and MCP server for lightweight, Chrome-free web discovery, with explicit escalation to Unchained for browser-only flows.
How do I install unbrowser?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is unbrowser open source?
Yes — it is hosted on GitHub at https://github.com/protostatis/unbrowser and has 15 stars.
Related MCP tools
Fast, local-first web content extraction for LLMs. Scrape, crawl, extract structured data — all from Rust. CLI, REST API, and MCP server.
an open source, extensible AI agent that goes beyond code suggestions - install, execute, edit, and test with any LLM
YC (S26) | Open Computer History | Record your screen continuously locally and provide context to your agents (Claude, Codex, Openclaw, Hermes, Runner...)
Markdown knowledge graph — LSP for your editor, CLI + MCP memory for your AI agents
🔥 Official Firecrawl MCP Server - Adds powerful web scraping and search to Cursor, Claude and any other LLM clients.
Browser automation CLI built for AI agents. Break through anti-bot walls, hand off to humans across platforms when stuck. Parallel multi-task execution, independent multi-session operation, isolated multi-account browsing.
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP