Agentic Browser

The web, as an LLM wants to see it

Stable refs · compact observations · outcome-verified actions · MCP for any agent host

Not another HTML scraper. A real Chromium browser controlled with semantic tools — so Claude, Cursor, LangGraph, or your own agent can see, click, type, and verify without drowning in 100k+ tokens of markup.

Repo: github.com/applejuice093/Agentic-browser · Branch: dev

---

Why this exists

Traditional stacks hand the model raw HTML or brittle CSS. Agents need:

1. Small, structured observations (roles, labels, refs)
2. Actions that mean success (URL/DOM outcomes, not “no exception”)
3. A way to plug into any host (MCP + OpenAI/Anthropic tool schemas)

Agentic Browser is built for that loop.

navigate → observe (~1–2k tokens) → click_text / type → outcome_verified?
         ↘ if page_gate = challenge → stop (don’t hallucinate)

---

Highlights — measured positives

Numbers from live benchmarks on public sites (approx. tokens ≈ chars/4). Reproduce with scripts under examples/.

Token efficiency (LLM feed)

Scenario	Raw HTML → model	Our compact observation / structured feed	Reduction
Quotes scrape (structured answers)	~2.8k–6.2k	~0.45k–1.3k	~78–84% fewer tokens
Rockstar GTA VI landing	~225,000	~1,300	~99.4% fewer tokens
GitHub vercel/next.js	~110,000	~1,900	~98.3% fewer tokens

Takeaway: Dumping HTML into an LLM is the wrong default. Observation mode is ~98–99% smaller on heavy modern pages.

Task quality

Metric	Traditional HTTP + BS4	Agentic Browser	Notes
Quote scrape field completeness	100% (with site CSS)	100%	Parity on friendly sites
Plain-text “LLM scrape” completeness	0% tags lost	100% via tools	Text dumps drop structure
Content signals on GitHub repo	~70%	~80%	Slight edge after JS + observe
Actionable element refs	0	20–90+ per page	Only we can click by ref
GitHub “open Issues” (after v0.3.2)	N/A	Verified → URL /issues	Outcome-checked, not false-ok
Network / XHR visibility	0%	Full request log + GraphQL flags	Agent can wait on APIs

Product completeness

Area	Coverage
Milestone roadmap M1–M10	100% of planned milestones delivered (plus agent-native v0.3–0.4)
Automated tests on dev	118 passed (full suite)
MCP tools for hosts	10 tools (navigate, observe, click, type, wait, find, network, …)
Observation token budget (default)	~2,000 (configurable)

Positives at a glance

• ~98–99% token cut vs raw HTML on complex landings (Rockstar, GitHub)
• ~78–84% token cut vs HTML for structured extract tasks
• 100% structured field parity with best-case CSS scrapers on demo sites
• Outcome verification — ok only when post-conditions hold (e.g. /issues in URL)
• Scoped grounding — nav-first find so PR/commit text doesn’t steal “Issues”
• Page gates — js_challenge / captcha / login wall detected and reported
• MCP-ready — Claude Desktop, Cursor, any MCP host
• OpenAI + Anthropic tool schemas — tools_as_openai() / tools_as_anthropic()
• Cookie/CMP dismiss, SPA settle budgets, stale-ref recovery, network intelligence
• Privacy defaults: mask Authorization / cookies in network logs

---

Bottlenecks & negatives (no sugarcoating)

We measure these so you don’t ship blind.

Issue	Reality	Impact
Latency	Real browser is ~6–16× slower than httpx on the same URL	Bad for high-QPS crawl; fine for agent steps
Bot walls	Reddit-class JS challenges: both HTTP and browser get ~17% content signals	Not a bypass tool — we surface page_gate and stop
SSR-only read	On some marketing pages, plain HTTP already has the copy	Agent wins on actions, not always on pure read speed
Full semantic snapshot	Can be larger than HTML if you dump everything	Always use observe(), not full tree, for LLMs
Speed vs design goal	Sub-500 ms action latency is not met end-to-end on heavy SPAs	Dominated by page load + settle, not Python
False success (historical)	Pre-0.3.2 GitHub “Issues” click could report ok without navigating	Fixed with outcome verification + GitHub skill
Skills coverage	GitHub tab skill is first-class; other domains need packs	Extensible under agent/skills/
Vision	OCR optional; not in the default observe loop	Canvas-heavy UIs still weaker

Honest product line:
Best as an LLM action + compact perception layer.
Not a replacement for bulk HTTP scraping, and not a captcha solver.

---

Architecture

┌──────────────────────────────────────────────────────┐
│  Claude · Cursor · LangGraph · OpenAI tool loop      │
└──────────────────────────┬───────────────────────────┘
                           │ MCP stdio  or  function tools
                           ▼
┌──────────────────────────────────────────────────────┐
│  agent-browser MCP  ·  tools_as_openai()             │
│  AgentSession: observe · click_text · wait · network │
└──────────────────────────┬───────────────────────────┘
                           ▼
┌──────────────────────────────────────────────────────┐
│  Semantic DOM · scoped grounding · outcomes · gates  │
│  Network monitor · memory · settle · overlay dismiss │
└──────────────────────────┬───────────────────────────┘
                           ▼
                 Playwright · Chromium

---

Installation

Requirements

• Python 3.11+
• Git
• ~Windows / macOS / Linux

Library + dev

git clone https://github.com/applejuice093/Agentic-browser.git
cd Agentic-browser
git checkout dev

python -m venv .venv

# Windows
.venv\Scripts\activate
# macOS / Linux
# source .venv/bin/activate

pip install -e ".[dev]"
playwright install chromium
pytest -q

With MCP (Claude / Cursor)

pip install -e ".[mcp,dev]"
playwright install chromium

Optional vision (local OCR)

pip install -e ".[vision]"
# Install system Tesseract; set TESSERACT_CMD on Windows

Extras summary

Extra	Install	Provides
default	pip install -e .	Core browser + agent API
mcp	pip install -e ".[mcp]"	MCP server for hosts
dev	pip install -e ".[dev]"	pytest, ruff, mypy
vision	pip install -e ".[vision]"	Pillow + pytesseract
all	pip install -e ".[all]"	Everything

---

Add to your agent (3 ways)

A) MCP — Claude Desktop / Cursor (recommended)

python -m agent_browser.mcp
# or
agent-browser-mcp

Cursor — .cursor/mcp.json (use your venv Python path):

{
  "mcpServers": {
    "agent-browser": {
      "command": "C:\\A\\PROJECT\\Agentic Browser\\.venv\\Scripts\\python.exe",
      "args": ["-m", "agent_browser.mcp"],
      "env": {
        "AGENT_BROWSER_HEADLESS": "true",
        "AGENT_BROWSER_MAX_TOKENS": "2000",
        "AGENT_BROWSER_ALLOWED_HOSTS": "github.com,example.com,quotes.toscrape.com"
      }
    }
  }
}

Claude Desktop — claude_desktop_config.json:

{
  "mcpServers": {
    "agent-browser": {
      "command": "C:\\A\\PROJECT\\Agentic Browser\\.venv\\Scripts\\python.exe",
      "args": ["-m", "agent_browser.mcp"],
      "env": { "AGENT_BROWSER_HEADLESS": "true" }
    }
  }
}

Full guide (env vars, system prompt, troubleshooting): docs/mcp.md

B) Python library

import asyncio
from agent_browser import Browser, tools_as_openai

async def main():
    async with Browser(headless=True) as browser:
        agent = await browser.open_agent("https://github.com/vercel/next.js")
        obs = await agent.observe(max_tokens=2000)
        if obs.page_gate not in (None, "open", "cookie_wall", "unknown"):
            print("Blocked:", obs.page_gate, obs.page_gate_hint)
            return
        result = await agent.click_text("Issues", scope="nav", intent="issues")
        print(result.ok, result.url_after, result.outcome_verified)

asyncio.run(main())

OpenAI / Anthropic tool schemas:

from agent_browser import tools_as_openai, tools_as_anthropic
tools = tools_as_openai()       # Chat Completions
# tools = tools_as_anthropic()  # Messages API

C) CLI

agent-browser version
agent-browser open https://example.com
agent-browser scrape https://example.com -o data/out.json

---

Core concepts

Concept	Meaning
ref	Stable integer id for click/type (from observe)
Observation	Compact LLM payload: interactive refs, headings, summary, gate, network hints
ActionResult	ok, error_code, outcome_verified, url_after, optional new observation
page_gate	open | js_challenge | captcha | login_wall | …
scope	nav / main / form — grounding region for find/click

MCP / tool surface

browser_navigate · browser_observe · browser_click · browser_click_text · browser_type · browser_wait · browser_find · browser_network · browser_resync · browser_prepare

---

Environment variables

Variable	Default	Purpose
AGENT_BROWSER_HEADLESS	true	Headless Chromium
AGENT_BROWSER_MAX_TOKENS	2000	Observation budget
AGENT_BROWSER_DETAIL	normal	sparse | normal | full
AGENT_BROWSER_SETTLE_MS	8000	SPA settle budget
AGENT_BROWSER_ALLOWED_HOSTS	(all)	Comma-separated allowlist
AGENT_BROWSER_DEFAULT_TIMEOUT_MS	30000	Action timeout
TESSERACT_CMD	—	OCR binary path (vision extra)

---

Project layout

src/agent_browser/
  browser.py / page.py / cli.py
  agent/          # AgentSession, MCP tools bridge, settle, overlays, grounding, outcomes, skills
  mcp/            # FastMCP stdio server
  observation/    # Compact observation builder
  semantic/ accessibility/ events/ network/ vision/ memory/ planning/
docs/
  mcp.md                 # Add to Claude / Cursor / custom agents
  USER_GUIDE.md
  agent-native-loop.md
  security.md
examples/
  agent_loop_demo.py
  openai_tool_agent.py
  mcp_config_cursor.json
  benchmark_hard_site.py
  compare_rockstar_vi.py
tests/                   # 118+ automated tests

---

Development

git checkout dev
pip install -e ".[dev,mcp]"
playwright install chromium
pytest -q
ruff check src tests

Doc	Topic
docs/mcp.md	Install into your agent (MCP + tools)
docs/USER_GUIDE.md	Full product guide
docs/agent-native-loop.md	Observe / act / gates / skills
docs/security.md	Security notes
CHANGELOG.md	Version history
deep-research-report.md	Original design research

---

Roadmap status

Track	Status
M1–M10 foundation	Done
Agent-native loop (v0.3)	Done
Overlays / settle / recovery (v0.3.1)	Done
Grounding + outcomes + GitHub skill (v0.3.2)	Done
MCP host integration (v0.4.0)	Done
More domain skills, vision-in-loop, pool/scale	Next

---

Ethics

• Use only on sites and accounts you’re allowed to automate.
• No captcha/JS-challenge bypass — we detect and report gates.
• Respect robots.txt / ToS / rate limits for your use case.
• Prefer AGENT_BROWSER_ALLOWED_HOSTS in shared demos.

---

License

MIT — see LICENSE.

---

Small observations. Real clicks. Verified outcomes.
_{Agentic Browser · built for LLMs that have to act on the web}