commit c1eef4c6aa6951a99b85fb375465ec600f2280e1
Author: Bhushan C Thakkar <hello@thecaoffice.com>
Date:   Mon Jul 27 07:43:33 2026 +0530

    Initial commit: Screen Leads app
    
    Screen-capture lead tool: capture agent, Claude vision extractor with
    pluggable site recipes (LinkedIn), canonical funnel model, SQLite storage,
    FastAPI backend, dashboard, and setup/run scripts.
    
    Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

diff --git a/.env.example b/.env.example
new file mode 100644
index 0000000..c234456
--- /dev/null
+++ b/.env.example
@@ -0,0 +1,20 @@
+# Copy to .env and fill in. Only ANTHROPIC_API_KEY is required.
+ANTHROPIC_API_KEY=sk-ant-...
+
+# Vision model (default is the most capable; sonnet is cheaper for high volume)
+# SCREEN_LEADS_MODEL=claude-opus-4-8
+
+# Capture tuning
+# SCREEN_LEADS_START_DELAY=5      # grace period to focus the browser (0 = off)
+# SCREEN_LEADS_SCROLL_STEPS=6
+# SCREEN_LEADS_SCROLL_AMOUNT=800
+# SCREEN_LEADS_SCROLL_PAUSE=0.8
+# SCREEN_LEADS_WATCH_INTERVAL=2.0
+
+# "stop" (default) ends the run when the front tab is not a LinkedIn profile;
+# "wait" keeps polling instead.
+# SCREEN_LEADS_ON_INVALID_PAGE=stop
+
+# Server
+# SCREEN_LEADS_HOST=127.0.0.1
+# SCREEN_LEADS_PORT=8000
diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..8234288
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,6 @@
+__pycache__/
+*.pyc
+.env
+data/
+.venv/
+venv/
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..2052dab
--- /dev/null
+++ b/README.md
@@ -0,0 +1,101 @@
+# Screen Leads
+
+Turn a browser tab you have open on a profile into a structured lead. The tool
+screenshots the screen, scrolls, runs the images through Claude vision, and
+saves normalised leads to a database you browse and control from a dashboard.
+
+Screen-capture only: it reads pixels a human already chose to display and never
+touches the target site's servers or DOM. LinkedIn is the first supported site;
+new sites drop in as recipes.
+
+## How it works
+
+```
+Start (dashboard) → detect front tab → [stop unless it's an enabled
+site's target page, e.g. a LinkedIn profile] → scroll + screenshot loop →
+Claude vision extraction → normalise to canonical Lead → SQLite → dashboard
+```
+
+- **Guard** — runs only when the front tab is an enabled recipe's *target* page
+  (a LinkedIn `/in/` profile). On anything else it stops (set
+  `SCREEN_LEADS_ON_INVALID_PAGE=wait` to poll instead).
+- **Detection** — reads your own browser's URL where possible (macOS reliably),
+  otherwise the AI reads the address bar from the screenshot. Both are passive.
+- **Contact info** — name/headline/company come from the profile page. Email &
+  phone live behind LinkedIn's **Contact info** panel: open it yourself before
+  capture and the tool will read whatever is shown. Often it simply isn't there.
+- **Funnel** — every lead is normalised to one schema and advances
+  `NEW → ENRICHED → CONTACT_FOUND → EXPORTED` as data completeness grows.
+
+## Setup
+
+```bash
+# 1. Get the code
+git clone https://git.thecaoffice.com/OpenSource/screen-leads-app.git
+cd screen-leads-app
+
+# 2. One-time setup — venv, dependencies, and prompts for your API key
+./setup.sh          # Windows: setup.bat
+
+# 3. Run it
+./run.sh            # Windows: run.bat
+```
+
+`setup.sh` creates the virtualenv, installs dependencies, and writes your
+`ANTHROPIC_API_KEY` to `.env`. `run.sh` launches the dashboard — open
+http://127.0.0.1:8000 and use **Start / Pause / Stop**.
+
+<details><summary>Manual steps (no scripts)</summary>
+
+```bash
+python3 -m venv .venv && source .venv/bin/activate
+pip install -r requirements.txt
+export ANTHROPIC_API_KEY=sk-ant-...
+python main.py
+```
+</details>
+
+### OS permissions (grant once)
+
+| OS | Needs |
+|----|-------|
+| macOS | System Settings → Privacy & Security → **Screen Recording** and **Accessibility** for your terminal/app. Automation permission for the browser enables exact URL detection. |
+| Windows | Usually none. `pip install uiautomation` enables native URL detection. |
+| Linux | Use an **X11** session. Wayland can't capture/scroll via `mss`/`pyautogui` — switch to X11 or use the desktop screenshot portal. |
+
+## Usage
+
+1. Open a LinkedIn profile in your normal browser (optionally open **Contact info**).
+2. Click **Start** on the dashboard.
+3. The tool scrolls, captures, extracts, and saves the lead; it then waits for
+   you to open the next profile. Navigating away from a profile stops the run
+   (default) so it never runs on non-target pages.
+
+## Project layout
+
+```
+screen-leads/
+  agent/          capture loop pieces: screenshot, scroll, detection
+    platform/     native URL helpers (mac / windows / linux)
+  ai/             Claude vision extraction + per-site recipes
+    recipes/      base.py, linkedin.py, registry (__init__.py)
+  core/           canonical models + capture controller (state machine)
+  db/             SQLite persistence (leads + runs)
+  api/            FastAPI: /api/start /pause /resume /stop /status /leads
+  dashboard/      single-file web UI
+  config.py  main.py
+```
+
+## Adding a new site
+
+1. Add `ai/recipes/<site>.py` implementing `SiteRecipe` (`matches`,
+   `extraction_schema`, `extraction_prompt`, `to_canonical`).
+2. Register it in `ai/recipes/__init__.py`.
+
+Nothing in the capture loop changes — the guard and pipeline are recipe-driven.
+
+## Compliance note
+
+Automated *scraping* of LinkedIn violates its Terms of Service. This tool is
+built for low-volume, human-in-the-loop use on profiles you manually open and
+are allowed to view. Keep pacing conservative and use it accordingly.
diff --git a/agent/__init__.py b/agent/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/agent/detector.py b/agent/detector.py
new file mode 100644
index 0000000..6a75ad7
--- /dev/null
+++ b/agent/detector.py
@@ -0,0 +1,70 @@
+"""Front-tab detection: decide whether the current page is an enabled site's
+capturable target (a LinkedIn profile, to start).
+
+Strategy:
+  1. Try to read the browser URL natively (fast, exact; macOS reliably, others
+     best-effort).
+  2. If no URL, capture one screenshot and let the AI read the address bar /
+     recognise the site visually (portable, screen-only).
+  3. Ask the recipe registry whether any enabled recipe matches.
+"""
+from __future__ import annotations
+
+from typing import Optional
+
+from agent import screenshot
+from agent.platform import get_active_tab_url
+from ai import recipes
+from core.models import DetectionResult, PageContext
+
+
+def detect(use_ai_fallback: bool = True) -> DetectionResult:
+    url = get_active_tab_url()
+    ctx = PageContext(url=url)
+    method = "url" if url else ""
+
+    hit = recipes.match(ctx)
+
+    if hit is None and url is None and use_ai_fallback:
+        # No native URL and nothing matched — read the screen visually.
+        try:
+            shot = screenshot.capture(prefix="detect")
+            from ai import extractor
+
+            info = extractor.classify_page(shot)
+            ctx.ai_site = info.get("site")
+            ctx.ai_page_type = info.get("page_type")
+            ctx.ai_url_text = info.get("address_bar_url")
+            method = "visual"
+            hit = recipes.match(ctx)
+        except Exception as e:
+            return DetectionResult(ok=False, reason=f"detection failed: {e}")
+
+    if hit is None:
+        where = ctx.best_url or "the current tab"
+        return DetectionResult(
+            ok=False,
+            reason=f"Front tab is not a supported site ({where}).",
+            url=ctx.best_url,
+            method=method,
+        )
+
+    recipe, pm = hit
+    if not pm.is_target:
+        return DetectionResult(
+            ok=False,
+            reason=(f"Front tab is {recipe.display_name} but not a "
+                    f"{recipe.target_page_type} page (got '{pm.page_type}')."),
+            recipe_id=recipe.site_id,
+            page_type=pm.page_type,
+            url=ctx.best_url,
+            method=method,
+        )
+
+    return DetectionResult(
+        ok=True,
+        recipe_id=recipe.site_id,
+        page_type=pm.page_type,
+        url=ctx.best_url,
+        method=method,
+    )
diff --git a/agent/platform/__init__.py b/agent/platform/__init__.py
new file mode 100644
index 0000000..51476af
--- /dev/null
+++ b/agent/platform/__init__.py
@@ -0,0 +1,30 @@
+"""Cross-platform helper for reading the front browser tab's URL.
+
+This is the one genuinely OS-specific piece. Reading *your own* browser's
+current URL is a passive, local operation (no request to the target site), so
+it stays within the screen-capture / compliance posture. When it can't return
+a URL (common on Windows/Linux), the detector falls back to reading the address
+bar visually from a screenshot via the AI.
+"""
+import sys
+from typing import Optional
+
+
+def get_active_tab_url() -> Optional[str]:
+    """Best-effort URL of the frontmost browser tab, or None if unavailable."""
+    try:
+        if sys.platform == "darwin":
+            from . import macos
+
+            return macos.get_active_tab_url()
+        if sys.platform.startswith("linux"):
+            from . import linux
+
+            return linux.get_active_tab_url()
+        if sys.platform in ("win32", "cygwin"):
+            from . import windows
+
+            return windows.get_active_tab_url()
+    except Exception:
+        return None
+    return None
diff --git a/agent/platform/linux.py b/agent/platform/linux.py
new file mode 100644
index 0000000..504c3a4
--- /dev/null
+++ b/agent/platform/linux.py
@@ -0,0 +1,11 @@
+"""Linux (X11): the active window *title* is reachable via xdotool, but the URL
+is not exposed without a browser extension or the accessibility bus. We return
+None so the detector falls back to reading the address bar visually from the
+screenshot, which is the portable path. (Wayland exposes neither; the app warns
+about that separately.)
+"""
+from typing import Optional
+
+
+def get_active_tab_url() -> Optional[str]:
+    return None
diff --git a/agent/platform/macos.py b/agent/platform/macos.py
new file mode 100644
index 0000000..1e5e16e
--- /dev/null
+++ b/agent/platform/macos.py
@@ -0,0 +1,39 @@
+"""macOS: read the frontmost browser tab URL via AppleScript.
+
+Reads local browser state only (no network call to the target site). Requires
+the terminal/app running this to have Automation permission for the browser.
+"""
+import subprocess
+from typing import Optional
+
+# Browsers that expose `URL of active tab of front window` (Chromium family)
+# or `URL of front document` (Safari).
+_CHROMIUM = ["Google Chrome", "Brave Browser", "Microsoft Edge", "Arc", "Vivaldi", "Chromium"]
+
+
+def _frontmost_app() -> Optional[str]:
+    script = 'tell application "System Events" to get name of first process whose frontmost is true'
+    return _osascript(script)
+
+
+def _osascript(script: str) -> Optional[str]:
+    try:
+        out = subprocess.run(
+            ["osascript", "-e", script],
+            capture_output=True, text=True, timeout=5,
+        )
+        val = out.stdout.strip()
+        return val or None
+    except Exception:
+        return None
+
+
+def get_active_tab_url() -> Optional[str]:
+    app = _frontmost_app()
+    if not app:
+        return None
+    if app in _CHROMIUM:
+        return _osascript(f'tell application "{app}" to get URL of active tab of front window')
+    if app == "Safari":
+        return _osascript('tell application "Safari" to get URL of front document')
+    return None
diff --git a/agent/platform/windows.py b/agent/platform/windows.py
new file mode 100644
index 0000000..cb5dd3c
--- /dev/null
+++ b/agent/platform/windows.py
@@ -0,0 +1,24 @@
+"""Windows: reliably reading a browser's address-bar URL needs UI Automation
+(pywinauto/uiautomation). We attempt a lightweight UIA read if available and
+otherwise return None, letting the detector fall back to visual address-bar
+reading from the screenshot.
+"""
+from typing import Optional
+
+
+def get_active_tab_url() -> Optional[str]:
+    try:
+        import uiautomation as auto  # optional dependency
+    except Exception:
+        return None
+    try:
+        window = auto.GetForegroundControl()
+        # Chromium/Firefox expose the address bar as an Edit control named
+        # "Address and search bar" / "Search or enter address".
+        edit = window.EditControl(searchDepth=12)
+        if edit.Exists(0.3):
+            val = (edit.GetValuePattern().Value or "").strip()
+            return val or None
+    except Exception:
+        return None
+    return None
diff --git a/agent/screenshot.py b/agent/screenshot.py
new file mode 100644
index 0000000..1907051
--- /dev/null
+++ b/agent/screenshot.py
@@ -0,0 +1,35 @@
+"""Cross-platform full-screen capture via `mss` (Windows, macOS, Linux/X11)."""
+from __future__ import annotations
+
+import time
+from pathlib import Path
+from typing import Optional
+
+from config import SCREENSHOT_DIR
+
+
+class CaptureError(RuntimeError):
+    pass
+
+
+def capture(prefix: str = "shot") -> str:
+    """Grab the primary monitor to a PNG and return its path."""
+    try:
+        import mss
+        import mss.tools
+    except Exception as e:  # pragma: no cover
+        raise CaptureError(f"mss not available: {e}")
+
+    ts = int(time.time() * 1000)
+    path = SCREENSHOT_DIR / f"{prefix}_{ts}.png"
+    try:
+        with mss.mss() as sct:
+            monitor = sct.monitors[1]  # [0] is the virtual "all monitors" rect
+            img = sct.grab(monitor)
+            mss.tools.to_png(img.rgb, img.size, output=str(path))
+    except Exception as e:
+        raise CaptureError(
+            f"Screen capture failed ({e}). On macOS grant Screen Recording; "
+            f"on Linux use an X11 session (Wayland needs the screenshot portal)."
+        )
+    return str(path)
diff --git a/agent/scroller.py b/agent/scroller.py
new file mode 100644
index 0000000..75931d2
--- /dev/null
+++ b/agent/scroller.py
@@ -0,0 +1,42 @@
+"""OS-level scrolling of the focused window via pyautogui.
+
+We drive real scroll input rather than the DOM, so it works over a plain
+screenshot pipeline on any OS. Needs Accessibility permission on macOS and an
+X11 session on Linux.
+"""
+from __future__ import annotations
+
+import time
+
+from config import SCROLL_AMOUNT, SCROLL_PAUSE
+
+
+class ScrollError(RuntimeError):
+    pass
+
+
+def _pyautogui():
+    try:
+        import pyautogui
+
+        pyautogui.FAILSAFE = False
+        return pyautogui
+    except Exception as e:  # pragma: no cover
+        raise ScrollError(
+            f"pyautogui not available ({e}). On macOS grant Accessibility; "
+            f"on Linux use X11."
+        )
+
+
+def scroll_to_top() -> None:
+    pg = _pyautogui()
+    # A large upward scroll reliably returns most pages to the top.
+    for _ in range(10):
+        pg.scroll(SCROLL_AMOUNT)
+    time.sleep(SCROLL_PAUSE)
+
+
+def scroll_down(amount: int = SCROLL_AMOUNT) -> None:
+    pg = _pyautogui()
+    pg.scroll(-abs(amount))
+    time.sleep(SCROLL_PAUSE)
diff --git a/ai/__init__.py b/ai/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/ai/extractor.py b/ai/extractor.py
new file mode 100644
index 0000000..09d5342
--- /dev/null
+++ b/ai/extractor.py
@@ -0,0 +1,90 @@
+"""Vision extraction via Claude.
+
+Two calls:
+  - extract_lead(): screenshots + a recipe's schema -> raw structured JSON.
+  - classify_page(): one screenshot -> {site, page_type, address bar url}, used
+    only when no native browser URL is available (visual site detection).
+
+Both use structured outputs (output_config.format) so the response is
+schema-validated JSON, not free text.
+"""
+from __future__ import annotations
+
+import base64
+import json
+from functools import lru_cache
+from typing import Any
+
+from config import ANTHROPIC_MODEL, MAX_IMAGES_PER_CALL
+from ai.recipes.base import SiteRecipe
+
+
+@lru_cache(maxsize=1)
+def _client():
+    import anthropic
+
+    # Resolves ANTHROPIC_API_KEY (or an `ant auth login` profile) from the env.
+    return anthropic.Anthropic()
+
+
+def _image_block(path: str) -> dict[str, Any]:
+    with open(path, "rb") as f:
+        data = base64.standard_b64encode(f.read()).decode("utf-8")
+    return {
+        "type": "image",
+        "source": {"type": "base64", "media_type": "image/png", "data": data},
+    }
+
+
+def _first_json(response) -> dict[str, Any]:
+    for block in response.content:
+        if block.type == "text":
+            return json.loads(block.text)
+    raise RuntimeError("model returned no text block")
+
+
+def extract_lead(recipe: SiteRecipe, screenshot_paths: list[str]) -> dict[str, Any]:
+    paths = screenshot_paths[:MAX_IMAGES_PER_CALL]
+    content: list[dict[str, Any]] = [_image_block(p) for p in paths]
+    content.append({"type": "text", "text": recipe.extraction_prompt})
+
+    response = _client().messages.create(
+        model=ANTHROPIC_MODEL,
+        max_tokens=2048,
+        messages=[{"role": "user", "content": content}],
+        output_config={"format": {"type": "json_schema", "schema": recipe.extraction_schema}},
+    )
+    return _first_json(response)
+
+
+_CLASSIFY_SCHEMA = {
+    "type": "object",
+    "properties": {
+        "site": {"type": "string"},        # e.g. "linkedin", "other"
+        "page_type": {"type": "string"},   # e.g. "profile", "search", "feed", "other"
+        "address_bar_url": {"type": "string"},
+    },
+    "required": ["site", "page_type", "address_bar_url"],
+    "additionalProperties": False,
+}
+
+_CLASSIFY_PROMPT = (
+    "This is a screenshot of a web browser. Identify the site and page from the "
+    "browser's address bar and the visible layout. Return: 'site' (lowercase, "
+    "e.g. 'linkedin', or 'other'); 'page_type' ('profile' for a single person's "
+    "page, otherwise 'search', 'feed', 'messaging', or 'other'); and "
+    "'address_bar_url' (the URL text you can read, or \"\")."
+)
+
+
+def classify_page(screenshot_path: str) -> dict[str, Any]:
+    response = _client().messages.create(
+        model=ANTHROPIC_MODEL,
+        max_tokens=512,
+        messages=[{
+            "role": "user",
+            "content": [_image_block(screenshot_path), {"type": "text", "text": _CLASSIFY_PROMPT}],
+        }],
+        output_config={"format": {"type": "json_schema", "schema": _CLASSIFY_SCHEMA}},
+    )
+    return _first_json(response)
diff --git a/ai/recipes/__init__.py b/ai/recipes/__init__.py
new file mode 100644
index 0000000..b5f44f4
--- /dev/null
+++ b/ai/recipes/__init__.py
@@ -0,0 +1,25 @@
+"""Recipe registry. Register new site recipes here — nothing else changes."""
+from __future__ import annotations
+
+from typing import Optional
+
+from core.models import PageContext, PageMatch
+from .base import SiteRecipe
+from .linkedin import LinkedInRecipe
+
+REGISTRY: list[SiteRecipe] = [
+    LinkedInRecipe(),
+]
+
+
+def enabled_recipes() -> list[SiteRecipe]:
+    return [r for r in REGISTRY if r.enabled]
+
+
+def match(ctx: PageContext) -> Optional[tuple[SiteRecipe, PageMatch]]:
+    """Return the first enabled recipe that recognises the page context."""
+    for recipe in enabled_recipes():
+        m = recipe.matches(ctx)
+        if m is not None:
+            return recipe, m
+    return None
diff --git a/ai/recipes/base.py b/ai/recipes/base.py
new file mode 100644
index 0000000..722206a
--- /dev/null
+++ b/ai/recipes/base.py
@@ -0,0 +1,36 @@
+"""SiteRecipe interface. Each supported site implements one of these; adding a
+new site is a drop-in module registered in ai/recipes/__init__.py — the core
+capture loop never changes.
+"""
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Any, Optional
+
+from core.models import Lead, PageContext, PageMatch
+
+
+class SiteRecipe(ABC):
+    site_id: str = ""
+    display_name: str = ""
+    enabled: bool = True
+    # Which extracted page_type is a capturable target (e.g. a person's profile).
+    target_page_type: str = "profile"
+
+    @abstractmethod
+    def matches(self, ctx: PageContext) -> Optional[PageMatch]:
+        """Return a PageMatch if this recipe recognises the page, else None."""
+
+    @property
+    @abstractmethod
+    def extraction_schema(self) -> dict[str, Any]:
+        """JSON schema (structured-output format) for the raw extraction."""
+
+    @property
+    @abstractmethod
+    def extraction_prompt(self) -> str:
+        """Instruction given to the vision model alongside the screenshots."""
+
+    @abstractmethod
+    def to_canonical(self, raw: dict[str, Any]) -> Lead:
+        """Map this site's raw extraction to the canonical Lead schema."""
diff --git a/ai/recipes/linkedin.py b/ai/recipes/linkedin.py
new file mode 100644
index 0000000..b5d5c42
--- /dev/null
+++ b/ai/recipes/linkedin.py
@@ -0,0 +1,91 @@
+"""LinkedIn profile recipe — the first supported site.
+
+Target page: a person's profile (linkedin.com/in/...), typically reached from
+search. Anything else on LinkedIn (feed, search results, messaging) is not a
+target; non-LinkedIn pages don't match at all.
+"""
+from __future__ import annotations
+
+import re
+from typing import Any, Optional
+
+from core.models import Lead, PageContext, PageMatch
+from .base import SiteRecipe
+
+_PROFILE_RE = re.compile(r"linkedin\.com/in/[^/?#]+", re.I)
+_LINKEDIN_RE = re.compile(r"linkedin\.com", re.I)
+
+
+class LinkedInRecipe(SiteRecipe):
+    site_id = "linkedin"
+    display_name = "LinkedIn"
+    enabled = True
+    target_page_type = "profile"
+
+    def matches(self, ctx: PageContext) -> Optional[PageMatch]:
+        url = ctx.best_url
+        if url:
+            if _PROFILE_RE.search(url):
+                return PageMatch(self.site_id, "profile", is_target=True)
+            if _LINKEDIN_RE.search(url):
+                return PageMatch(self.site_id, "other", is_target=False)
+            return None
+        # No URL available: fall back to the AI's visual read.
+        if (ctx.ai_site or "").lower() == "linkedin":
+            page_type = (ctx.ai_page_type or "other").lower()
+            return PageMatch(self.site_id, page_type, is_target=(page_type == "profile"))
+        return None
+
+    @property
+    def extraction_schema(self) -> dict[str, Any]:
+        str_fields = [
+            "full_name", "headline", "current_title", "current_company",
+            "location", "about", "email", "phone", "website",
+        ]
+        props: dict[str, Any] = {f: {"type": "string"} for f in str_fields}
+        props["other_profiles"] = {"type": "array", "items": {"type": "string"}}
+        props["confidence"] = {"type": "number"}
+        return {
+            "type": "object",
+            "properties": props,
+            "required": str_fields + ["other_profiles", "confidence"],
+            "additionalProperties": False,
+        }
+
+    @property
+    def extraction_prompt(self) -> str:
+        return (
+            "These screenshots are a single LinkedIn profile page, captured top "
+            "to bottom. Extract the person's details into the required JSON. "
+            "Rules: use an empty string \"\" for any field not visible; do NOT "
+            "guess or invent contact details. Email, phone and website usually "
+            "only appear if the user opened the 'Contact info' panel — extract "
+            "them only if they are actually shown. 'other_profiles' is any other "
+            "social/web links visible (Twitter/X, GitHub, personal site). "
+            "'confidence' is 0-1 for how legible the profile was."
+        )
+
+    def to_canonical(self, raw: dict[str, Any]) -> Lead:
+        socials: dict[str, str] = {}
+        for link in raw.get("other_profiles") or []:
+            low = link.lower()
+            if "github.com" in low:
+                socials["github"] = link
+            elif "twitter.com" in low or "x.com" in low:
+                socials["twitter"] = link
+            else:
+                socials.setdefault("other", link)
+        return Lead(
+            full_name=raw.get("full_name", "") or "",
+            headline=raw.get("headline", "") or "",
+            title=raw.get("current_title", "") or "",
+            company=raw.get("current_company", "") or "",
+            location=raw.get("location", "") or "",
+            about=raw.get("about", "") or "",
+            email=raw.get("email", "") or "",
+            phone=raw.get("phone", "") or "",
+            website=raw.get("website", "") or "",
+            social_handles=socials,
+            source_site=self.site_id,
+            confidence=float(raw.get("confidence", 0) or 0),
+        )
diff --git a/api/__init__.py b/api/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/api/app.py b/api/app.py
new file mode 100644
index 0000000..3621ed8
--- /dev/null
+++ b/api/app.py
@@ -0,0 +1,72 @@
+"""FastAPI backend: run controls, status, and lead queries; serves the dashboard."""
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Optional
+
+from fastapi import FastAPI
+from fastapi.responses import HTMLResponse, JSONResponse
+
+import config
+from core.controller import controller
+from db import database
+
+app = FastAPI(title="Screen Leads")
+
+_ROOT = Path(__file__).resolve().parent.parent
+_DASHBOARD_DIR = _ROOT / "dashboard"
+_DASHBOARD = _DASHBOARD_DIR / "index.html"
+_SUPPORT = _DASHBOARD_DIR / "support.html"
+
+
+@app.on_event("startup")
+def _startup() -> None:
+    database.init_db()
+
+
+@app.get("/", response_class=HTMLResponse)
+def index() -> str:
+    return _DASHBOARD.read_text(encoding="utf-8")
+
+
+@app.get("/support", response_class=HTMLResponse)
+def support() -> str:
+    return _SUPPORT.read_text(encoding="utf-8")
+
+
+# --- run controls ------------------------------------------------------------
+@app.post("/api/start")
+def start():
+    return controller.start()
+
+
+@app.post("/api/pause")
+def pause():
+    return controller.pause()
+
+
+@app.post("/api/resume")
+def resume():
+    return controller.resume()
+
+
+@app.post("/api/stop")
+def stop():
+    return controller.stop()
+
+
+@app.get("/api/status")
+def status():
+    return {**controller.status(), "stage_counts": database.stage_counts()}
+
+
+# --- leads -------------------------------------------------------------------
+@app.get("/api/leads")
+def leads(stage: Optional[str] = None):
+    return JSONResponse(database.list_leads(stage))
+
+
+@app.post("/api/leads/{lead_id}/stage")
+def set_stage(lead_id: int, stage: str):
+    database.set_stage(lead_id, stage)
+    return {"ok": True}
diff --git a/config.py b/config.py
new file mode 100644
index 0000000..b19bdc1
--- /dev/null
+++ b/config.py
@@ -0,0 +1,55 @@
+"""Central configuration, all overridable via environment variables.
+
+Anything user- or machine-specific lives here so the rest of the code never
+reads os.environ directly.
+"""
+import os
+from pathlib import Path
+
+try:
+    from dotenv import load_dotenv
+
+    load_dotenv()
+except Exception:  # dotenv is optional
+    pass
+
+
+def _env(name: str, default: str) -> str:
+    return os.environ.get(name, default)
+
+
+BASE_DIR = Path(__file__).resolve().parent
+
+# --- Storage -----------------------------------------------------------------
+DATA_DIR = Path(_env("SCREEN_LEADS_DATA", str(BASE_DIR / "data")))
+DATA_DIR.mkdir(parents=True, exist_ok=True)
+
+SCREENSHOT_DIR = DATA_DIR / "screenshots"
+SCREENSHOT_DIR.mkdir(parents=True, exist_ok=True)
+
+DB_PATH = Path(_env("SCREEN_LEADS_DB", str(DATA_DIR / "leads.db")))
+
+# --- AI ----------------------------------------------------------------------
+# Vision model used for extraction. Defaults to the most capable Claude model;
+# override with SCREEN_LEADS_MODEL=claude-sonnet-5 for a cheaper high-volume run.
+ANTHROPIC_MODEL = _env("SCREEN_LEADS_MODEL", "claude-opus-4-8")
+MAX_IMAGES_PER_CALL = int(_env("SCREEN_LEADS_MAX_IMAGES", "8"))
+
+# --- Capture -----------------------------------------------------------------
+# Grace period after Start before the first capture, so you can focus the
+# browser window on the profile. Set to 0 to disable.
+START_DELAY = float(_env("SCREEN_LEADS_START_DELAY", "5"))
+
+SCROLL_STEPS = int(_env("SCREEN_LEADS_SCROLL_STEPS", "6"))   # screenshots per profile
+SCROLL_AMOUNT = int(_env("SCREEN_LEADS_SCROLL_AMOUNT", "800"))  # pyautogui scroll clicks
+SCROLL_PAUSE = float(_env("SCREEN_LEADS_SCROLL_PAUSE", "0.8"))  # settle time (s)
+WATCH_INTERVAL = float(_env("SCREEN_LEADS_WATCH_INTERVAL", "2.0"))  # loop tick (s)
+
+# What to do when the front tab is not an enabled site's target page.
+#   "stop" -> end the run (the requested behaviour)
+#   "wait" -> pause and keep polling until a valid page appears
+ON_INVALID_PAGE = _env("SCREEN_LEADS_ON_INVALID_PAGE", "stop").lower()
+
+# --- Server ------------------------------------------------------------------
+HOST = _env("SCREEN_LEADS_HOST", "127.0.0.1")
+PORT = int(_env("SCREEN_LEADS_PORT", "8000"))
diff --git a/core/__init__.py b/core/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/core/controller.py b/core/controller.py
new file mode 100644
index 0000000..5eb7ed2
--- /dev/null
+++ b/core/controller.py
@@ -0,0 +1,186 @@
+"""Capture controller: a background state machine driven by the dashboard.
+
+States: IDLE -> RUNNING <-> PAUSED -> STOPPED. The loop detects the front tab,
+stops (or waits) on a non-target page per config, and for each new profile URL
+runs scroll -> screenshot -> extract -> normalise -> save.
+"""
+from __future__ import annotations
+
+import threading
+import time
+import traceback
+from enum import Enum
+from typing import Any, Optional
+
+import config
+from agent import detector, screenshot, scroller
+from ai import extractor, recipes
+from core.models import Lead
+from db import database
+
+
+class State(str, Enum):
+    IDLE = "IDLE"
+    RUNNING = "RUNNING"
+    PAUSED = "PAUSED"
+    STOPPED = "STOPPED"
+
+
+def _recipe_by_id(site_id: str):
+    for r in recipes.REGISTRY:
+        if r.site_id == site_id:
+            return r
+    return None
+
+
+class CaptureController:
+    def __init__(self) -> None:
+        self._state = State.IDLE
+        self._thread: Optional[threading.Thread] = None
+        self._stop = threading.Event()
+        self._pause = threading.Event()
+        self._lock = threading.Lock()
+        self.reason = ""
+        self.last_event = ""
+        self.captured_urls: set[str] = set()
+        self.leads_captured = 0
+        self._run_id: Optional[int] = None
+
+    # --- introspection -------------------------------------------------------
+    def status(self) -> dict[str, Any]:
+        return {
+            "state": self._state.value,
+            "reason": self.reason,
+            "last_event": self.last_event,
+            "leads_captured": self.leads_captured,
+            "captured_urls": len(self.captured_urls),
+            "on_invalid_page": config.ON_INVALID_PAGE,
+        }
+
+    def _set(self, state: State, event: str = "") -> None:
+        self._state = state
+        if event:
+            self.last_event = event
+
+    # --- controls ------------------------------------------------------------
+    def start(self) -> dict[str, Any]:
+        with self._lock:
+            if self._state in (State.RUNNING, State.PAUSED):
+                return self.status()
+            self._stop.clear()
+            self._pause.clear()
+            self.reason = ""
+            self.leads_captured = 0
+            self.captured_urls.clear()
+            self._run_id = database.start_run()
+            self._set(State.RUNNING, "started")
+            self._thread = threading.Thread(target=self._loop, daemon=True)
+            self._thread.start()
+        return self.status()
+
+    def pause(self) -> dict[str, Any]:
+        if self._state == State.RUNNING:
+            self._pause.set()
+            self._set(State.PAUSED, "paused")
+        return self.status()
+
+    def resume(self) -> dict[str, Any]:
+        if self._state == State.PAUSED:
+            self._pause.clear()
+            self._set(State.RUNNING, "resumed")
+        return self.status()
+
+    def stop(self) -> dict[str, Any]:
+        if self._state in (State.RUNNING, State.PAUSED):
+            self._stop.set()
+            self._pause.clear()
+            self.reason = self.reason or "stopped by user"
+        return self.status()
+
+    # --- main loop -----------------------------------------------------------
+    def _finish(self, status: str, reason: str) -> None:
+        self.reason = reason
+        self._set(State.STOPPED, reason)
+        if self._run_id is not None:
+            database.finish_run(self._run_id, status, reason, self.leads_captured)
+            self._run_id = None
+
+    def _countdown(self) -> None:
+        """Grace period so the user can focus the browser before capture."""
+        remaining = int(round(config.START_DELAY))
+        while remaining > 0 and not self._stop.is_set():
+            self.last_event = f"focus your browser — starting in {remaining}s"
+            time.sleep(1)
+            remaining -= 1
+
+    def _loop(self) -> None:
+        try:
+            self._countdown()
+            while not self._stop.is_set():
+                if self._pause.is_set():
+                    time.sleep(0.3)
+                    continue
+
+                det = detector.detect()
+
+                if not det.ok:
+                    if config.ON_INVALID_PAGE == "wait":
+                        self.last_event = f"waiting: {det.reason}"
+                        time.sleep(config.WATCH_INTERVAL)
+                        continue
+                    self._finish("stopped", det.reason)
+                    return
+
+                url = det.url or ""
+                if url and url in self.captured_urls:
+                    self.last_event = "waiting for a new profile"
+                    time.sleep(config.WATCH_INTERVAL)
+                    continue
+
+                recipe = _recipe_by_id(det.recipe_id or "")
+                if recipe is None:
+                    self._finish("error", f"no recipe for {det.recipe_id}")
+                    return
+
+                self.last_event = f"capturing {url or det.recipe_id}"
+                lead = self._capture(recipe, url)
+                if lead is not None:
+                    self.leads_captured += 1
+                    if url:
+                        self.captured_urls.add(url)
+                    self.last_event = f"saved: {lead.full_name or url}"
+
+                time.sleep(config.WATCH_INTERVAL)
+
+            self._finish("stopped", self.reason or "stopped by user")
+        except Exception as e:  # keep the loop crash-visible on the dashboard
+            self._finish("error", f"{e}\n{traceback.format_exc(limit=2)}")
+
+    def _capture(self, recipe, url: str) -> Optional[Lead]:
+        # scroll + screenshot the full page
+        try:
+            scroller.scroll_to_top()
+        except Exception as e:
+            self.last_event = f"scroll unavailable: {e}"
+        shots: list[str] = []
+        for i in range(config.SCROLL_STEPS):
+            if self._stop.is_set():
+                break
+            shots.append(screenshot.capture(prefix=f"{recipe.site_id}"))
+            try:
+                scroller.scroll_down()
+            except Exception:
+                break
+        if not shots:
+            return None
+
+        raw = extractor.extract_lead(recipe, shots)
+        lead = recipe.to_canonical(raw)
+        lead.source_url = url
+        lead.screenshot_refs = shots
+        database.upsert_lead(lead)
+        return lead
+
+
+# module-level singleton shared by the API
+controller = CaptureController()
diff --git a/core/models.py b/core/models.py
new file mode 100644
index 0000000..57f3708
--- /dev/null
+++ b/core/models.py
@@ -0,0 +1,115 @@
+"""Canonical, site-agnostic data structures.
+
+Every site recipe normalises its raw extraction into `Lead`, so leads from any
+source share one funnel table and one dashboard view.
+"""
+from __future__ import annotations
+
+import json
+import time
+from dataclasses import dataclass, field, asdict
+from enum import Enum
+from typing import Any, Optional
+
+
+class FunnelStage(str, Enum):
+    NEW = "NEW"                    # name/headline only
+    ENRICHED = "ENRICHED"          # company / title / location / about captured
+    CONTACT_FOUND = "CONTACT_FOUND"  # email / phone / website captured
+    EXPORTED = "EXPORTED"          # pushed downstream (set by the user)
+
+
+@dataclass
+class Lead:
+    # identity
+    full_name: str = ""
+    first_name: str = ""
+    last_name: str = ""
+    headline: str = ""
+    title: str = ""
+    # organisation
+    company: str = ""
+    company_url: str = ""
+    industry: str = ""
+    location: str = ""
+    about: str = ""
+    # contact (often empty on the first pass)
+    email: str = ""
+    phone: str = ""
+    website: str = ""
+    social_handles: dict[str, str] = field(default_factory=dict)
+    # provenance
+    source_site: str = ""
+    source_url: str = ""
+    captured_at: float = field(default_factory=time.time)
+    updated_at: float = field(default_factory=time.time)
+    screenshot_refs: list[str] = field(default_factory=list)
+    # quality
+    confidence: float = 0.0
+    fields_found: list[str] = field(default_factory=list)
+    needs_review: bool = False
+    # funnel
+    stage: str = FunnelStage.NEW.value
+
+    def compute_stage(self) -> str:
+        """Advance the funnel stage from data completeness (never regress EXPORTED)."""
+        if self.stage == FunnelStage.EXPORTED.value:
+            return self.stage
+        if self.email or self.phone or self.website or self.social_handles:
+            return FunnelStage.CONTACT_FOUND.value
+        if self.company or self.title or self.location or self.about:
+            return FunnelStage.ENRICHED.value
+        return FunnelStage.NEW.value
+
+    def finalize(self) -> "Lead":
+        """Derive first/last name, fields_found, and funnel stage."""
+        if self.full_name and not (self.first_name or self.last_name):
+            parts = self.full_name.split()
+            if parts:
+                self.first_name = parts[0]
+                self.last_name = " ".join(parts[1:])
+        self.fields_found = [
+            k for k, v in asdict(self).items()
+            if k not in ("fields_found", "stage", "needs_review", "captured_at",
+                         "updated_at", "confidence") and v
+        ]
+        self.stage = self.compute_stage()
+        return self
+
+    def to_dict(self) -> dict[str, Any]:
+        return asdict(self)
+
+
+@dataclass
+class PageMatch:
+    """Result of a recipe testing a page context."""
+    site_id: str
+    page_type: str          # e.g. "profile", "search", "feed", "other"
+    is_target: bool         # True only for a page we should capture
+
+
+@dataclass
+class PageContext:
+    """What the detector knows about the current front tab."""
+    url: Optional[str] = None
+    ai_site: Optional[str] = None       # site guessed visually by the AI
+    ai_page_type: Optional[str] = None  # page type guessed visually
+    ai_url_text: Optional[str] = None   # address-bar text read from a screenshot
+
+    @property
+    def best_url(self) -> str:
+        return (self.url or self.ai_url_text or "").strip()
+
+
+@dataclass
+class DetectionResult:
+    ok: bool                       # a valid, enabled target page is in front
+    reason: str = ""               # human-readable explanation (shown on stop)
+    recipe_id: Optional[str] = None
+    page_type: Optional[str] = None
+    url: str = ""
+    method: str = ""               # "url" | "visual"
+
+
+def dumps(obj: Any) -> str:
+    return json.dumps(obj, ensure_ascii=False)
diff --git a/dashboard/index.html b/dashboard/index.html
new file mode 100644
index 0000000..05ab8aa
--- /dev/null
+++ b/dashboard/index.html
@@ -0,0 +1,121 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1" />
+<title>Screen Leads</title>
+<style>
+  :root {
+    --bg: #0f1115; --panel: #171a21; --border: #262b36; --fg: #e6e9ef;
+    --muted: #8b93a3; --accent: #4f8cff; --ok: #34c759; --warn: #ffb020; --danger: #ff5c5c;
+  }
+  @media (prefers-color-scheme: light) {
+    :root { --bg:#f5f6f8; --panel:#fff; --border:#e2e5ea; --fg:#1b1f27; --muted:#6b7280; }
+  }
+  * { box-sizing: border-box; }
+  body { margin:0; font:14px/1.5 -apple-system,Segoe UI,Roboto,Helvetica,Arial,sans-serif;
+         background:var(--bg); color:var(--fg); }
+  header { display:flex; align-items:center; gap:12px; padding:14px 20px;
+           border-bottom:1px solid var(--border); background:var(--panel); position:sticky; top:0; }
+  header h1 { font-size:16px; margin:0; font-weight:600; }
+  header a { color:var(--accent); text-decoration:none; font-size:13px; font-weight:600; }
+  header a:hover { text-decoration:underline; }
+  .pill { padding:2px 10px; border-radius:999px; font-size:12px; font-weight:600; border:1px solid var(--border); }
+  .s-IDLE{color:var(--muted)} .s-RUNNING{color:var(--ok);border-color:var(--ok)}
+  .s-PAUSED{color:var(--warn);border-color:var(--warn)} .s-STOPPED{color:var(--danger);border-color:var(--danger)}
+  main { padding:20px; max-width:1200px; margin:0 auto; }
+  .controls { display:flex; gap:8px; flex-wrap:wrap; margin-left:auto; }
+  button { font:inherit; font-weight:600; padding:7px 14px; border-radius:8px; cursor:pointer;
+           border:1px solid var(--border); background:var(--panel); color:var(--fg); }
+  button:hover { border-color:var(--accent); }
+  button:disabled { opacity:.4; cursor:not-allowed; }
+  button.primary { background:var(--accent); border-color:var(--accent); color:#fff; }
+  .event { color:var(--muted); font-size:12px; padding:8px 20px; border-bottom:1px solid var(--border);
+           background:var(--panel); }
+  .stages { display:flex; gap:10px; flex-wrap:wrap; margin-bottom:16px; }
+  .stage { background:var(--panel); border:1px solid var(--border); border-radius:10px; padding:10px 14px; min-width:120px; }
+  .stage .n { font-size:22px; font-weight:700; }
+  .stage .l { font-size:11px; color:var(--muted); letter-spacing:.04em; }
+  table { width:100%; border-collapse:collapse; background:var(--panel);
+          border:1px solid var(--border); border-radius:10px; overflow:hidden; }
+  th,td { text-align:left; padding:9px 12px; border-bottom:1px solid var(--border); vertical-align:top; }
+  th { font-size:11px; letter-spacing:.05em; color:var(--muted); text-transform:uppercase; }
+  tr:last-child td { border-bottom:0; }
+  .tag { font-size:11px; padding:1px 8px; border-radius:6px; border:1px solid var(--border); white-space:nowrap; }
+  .tag-NEW{color:var(--muted)} .tag-ENRICHED{color:var(--accent);border-color:var(--accent)}
+  .tag-CONTACT_FOUND{color:var(--ok);border-color:var(--ok)} .tag-EXPORTED{color:var(--warn);border-color:var(--warn)}
+  a { color:var(--accent); text-decoration:none; }
+  .muted { color:var(--muted); }
+  .empty { text-align:center; color:var(--muted); padding:40px; }
+</style>
+</head>
+<body>
+<header>
+  <h1>Screen&nbsp;Leads</h1>
+  <span id="state" class="pill s-IDLE">IDLE</span>
+  <div class="controls">
+    <button id="btn-start" class="primary" onclick="act('start')">Start</button>
+    <button id="btn-pause" onclick="act('pause')">Pause</button>
+    <button id="btn-resume" onclick="act('resume')">Resume</button>
+    <button id="btn-stop" onclick="act('stop')">Stop</button>
+    <a href="/support">Help</a>
+  </div>
+</header>
+<div id="event" class="event">—</div>
+<main>
+  <div id="stages" class="stages"></div>
+  <table>
+    <thead><tr>
+      <th>Name</th><th>Title / Company</th><th>Location</th><th>Contact</th>
+      <th>Stage</th><th>Conf.</th><th>Captured</th><th>Source</th>
+    </tr></thead>
+    <tbody id="rows"><tr><td colspan="8" class="empty">No leads yet.</td></tr></tbody>
+  </table>
+</main>
+<script>
+const STAGES = ["NEW","ENRICHED","CONTACT_FOUND","EXPORTED"];
+async function api(path, method="GET"){ const r=await fetch(path,{method}); return r.json(); }
+async function act(name){ await api("/api/"+name,"POST"); refresh(); }
+function esc(s){ return (s||"").replace(/[&<>]/g,c=>({"&":"&amp;","<":"&lt;",">":"&gt;"}[c])); }
+function fmtDate(ts){ if(!ts) return '<span class="muted">—</span>';
+  const d=new Date(ts*1000);
+  return d.toLocaleDateString(undefined,{year:'numeric',month:'short',day:'numeric'})
+    +'<br><span class="muted">'+d.toLocaleTimeString(undefined,{hour:'2-digit',minute:'2-digit'})+'</span>'; }
+
+async function refresh(){
+  const st = await api("/api/status");
+  const el = document.getElementById("state");
+  el.textContent = st.state; el.className = "pill s-"+st.state;
+  document.getElementById("event").textContent =
+    (st.last_event||"—") + (st.reason && st.state==="STOPPED" ? "  ·  "+st.reason : "");
+  document.getElementById("btn-start").disabled  = st.state==="RUNNING"||st.state==="PAUSED";
+  document.getElementById("btn-pause").disabled  = st.state!=="RUNNING";
+  document.getElementById("btn-resume").disabled = st.state!=="PAUSED";
+  document.getElementById("btn-stop").disabled   = st.state!=="RUNNING"&&st.state!=="PAUSED";
+
+  const counts = st.stage_counts||{};
+  document.getElementById("stages").innerHTML = STAGES.map(s=>
+    `<div class="stage"><div class="n">${counts[s]||0}</div><div class="l">${s.replace("_"," ")}</div></div>`).join("");
+
+  const leads = await api("/api/leads");
+  const rows = document.getElementById("rows");
+  if(!leads.length){ rows.innerHTML='<tr><td colspan="7" class="empty">No leads yet.</td></tr>'; return; }
+  rows.innerHTML = leads.map(l=>{
+    const contact=[l.email,l.phone,l.website].filter(Boolean).map(esc).join("<br>")||'<span class="muted">—</span>';
+    const src = l.source_url?`<a href="${esc(l.source_url)}" target="_blank" rel="noopener">${esc(l.source_site||"link")}</a>`:esc(l.source_site);
+    return `<tr>
+      <td><strong>${esc(l.full_name)||'<span class="muted">?</span>'}</strong><br><span class="muted">${esc(l.headline)}</span></td>
+      <td>${esc(l.title)}${l.company?' · '+esc(l.company):''}</td>
+      <td>${esc(l.location)}</td>
+      <td>${contact}</td>
+      <td><span class="tag tag-${l.stage}">${l.stage.replace("_"," ")}</span></td>
+      <td>${l.confidence?Math.round(l.confidence*100)+'%':'—'}</td>
+      <td>${fmtDate(l.captured_at)}</td>
+      <td>${src}</td>
+    </tr>`;
+  }).join("");
+}
+refresh(); setInterval(refresh, 2000);
+</script>
+</body>
+</html>
diff --git a/dashboard/support.html b/dashboard/support.html
new file mode 100644
index 0000000..2b99839
--- /dev/null
+++ b/dashboard/support.html
@@ -0,0 +1,173 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8" />
+<meta name="viewport" content="width=device-width, initial-scale=1" />
+<title>Screen Leads · Support</title>
+<style>
+  :root {
+    --bg: #0f1115; --panel: #171a21; --border: #262b36; --fg: #e6e9ef;
+    --muted: #8b93a3; --accent: #4f8cff; --ok: #34c759; --warn: #ffb020; --danger: #ff5c5c;
+  }
+  @media (prefers-color-scheme: light) {
+    :root { --bg:#f5f6f8; --panel:#fff; --border:#e2e5ea; --fg:#1b1f27; --muted:#6b7280; }
+  }
+  * { box-sizing: border-box; }
+  body { margin:0; font:15px/1.65 -apple-system,Segoe UI,Roboto,Helvetica,Arial,sans-serif;
+         background:var(--bg); color:var(--fg); }
+  header { display:flex; align-items:center; gap:12px; padding:14px 20px;
+           border-bottom:1px solid var(--border); background:var(--panel); position:sticky; top:0; }
+  header h1 { font-size:16px; margin:0; font-weight:600; }
+  header a { margin-left:auto; }
+  a { color:var(--accent); text-decoration:none; }
+  a:hover { text-decoration:underline; }
+  main { padding:24px 20px 60px; max-width:820px; margin:0 auto; }
+  h2 { font-size:18px; margin:34px 0 10px; padding-bottom:6px; border-bottom:1px solid var(--border); }
+  h3 { font-size:15px; margin:20px 0 6px; }
+  p, li { color:var(--fg); }
+  .muted { color:var(--muted); }
+  code, pre { font-family:ui-monospace,SFMono-Regular,Menlo,monospace; font-size:13px; }
+  code { background:var(--panel); border:1px solid var(--border); border-radius:5px; padding:1px 6px; }
+  pre { background:var(--panel); border:1px solid var(--border); border-radius:10px; padding:14px 16px;
+        overflow-x:auto; }
+  pre code { border:0; background:none; padding:0; }
+  table { width:100%; border-collapse:collapse; margin:12px 0; background:var(--panel);
+          border:1px solid var(--border); border-radius:10px; overflow:hidden; }
+  th,td { text-align:left; padding:9px 12px; border-bottom:1px solid var(--border); vertical-align:top; }
+  th { font-size:12px; color:var(--muted); text-transform:uppercase; letter-spacing:.04em; }
+  tr:last-child td { border-bottom:0; }
+  .toc { display:flex; flex-wrap:wrap; gap:8px 16px; margin:6px 0 10px; }
+  .note { border-left:3px solid var(--accent); background:var(--panel); border-radius:0 8px 8px 0;
+          padding:10px 14px; margin:14px 0; }
+  .warn { border-left-color:var(--warn); }
+</style>
+</head>
+<body>
+<header>
+  <h1>Screen&nbsp;Leads · Support</h1>
+  <a href="/">← Back to dashboard</a>
+</header>
+<main>
+  <p class="muted">Everything you need to install, run, and troubleshoot the tool.</p>
+  <div class="toc">
+    <a href="#what">What it does</a>
+    <a href="#start">Quick start</a>
+    <a href="#perms">Permissions</a>
+    <a href="#use">Using it</a>
+    <a href="#controls">Controls</a>
+    <a href="#trouble">Troubleshooting</a>
+    <a href="#config">Configuration</a>
+    <a href="#sites">Adding a site</a>
+    <a href="#compliance">Compliance</a>
+  </div>
+
+  <h2 id="what">What it does</h2>
+  <p>Screen Leads turns a browser tab you have open on a profile into a structured
+  lead. It screenshots your screen, scrolls, runs the images through Claude vision,
+  and saves normalised leads to a local database you browse from the dashboard.</p>
+  <p>It is <strong>screen-capture only</strong> — it reads pixels already on your
+  screen and never contacts the target site's servers. LinkedIn profile pages are
+  the first supported site.</p>
+
+  <h2 id="start">Quick start</h2>
+  <p>Run it on your own machine, in a terminal:</p>
+  <pre><code>cd path/to/screen-leads
+python3 -m venv .venv
+source .venv/bin/activate
+pip install -r requirements.txt
+export ANTHROPIC_API_KEY=sk-ant-...
+python main.py</code></pre>
+  <p>Then open <a href="http://127.0.0.1:8000">http://127.0.0.1:8000</a>. Next time,
+  skip the install steps — just activate the venv, set the key, and run
+  <code>python main.py</code>.</p>
+  <div class="note">Needs <code>ANTHROPIC_API_KEY</code> in the environment (or an
+  <code>ant auth login</code> profile). The key is only used to call Claude for extraction.</div>
+
+  <h2 id="perms">Permissions by OS</h2>
+  <table>
+    <tr><th>OS</th><th>Grant this</th></tr>
+    <tr><td>macOS</td><td>System Settings → Privacy &amp; Security → <strong>Screen Recording</strong>
+      and <strong>Accessibility</strong> for your terminal/app. Quit and reopen the terminal after
+      granting. Automation permission for the browser enables exact URL detection.</td></tr>
+    <tr><td>Windows</td><td>Usually none. <code>pip install uiautomation</code> enables native URL detection.</td></tr>
+    <tr><td>Linux</td><td>Use an <strong>X11</strong> session. Wayland can't capture or scroll via
+      <code>mss</code>/<code>pyautogui</code> — switch to X11 or use the desktop screenshot portal.</td></tr>
+  </table>
+
+  <h2 id="use">Using it</h2>
+  <ol>
+    <li>Open a LinkedIn profile in your normal browser. To capture email/phone, open
+      the <strong>Contact info</strong> panel first — those only appear there, and often
+      not at all.</li>
+    <li>Click <strong>Start</strong>. A short countdown appears so you can focus the browser window.</li>
+    <li>The tool scrolls, screenshots, extracts, and saves the lead, then waits for you
+      to open the next profile.</li>
+  </ol>
+  <div class="note warn">The tool captures the <strong>frontmost window</strong>. Keep the profile
+  tab focused and on top while it works — clicking back into the dashboard mid-capture points the
+  screenshots at the wrong window.</div>
+
+  <h2 id="controls">Dashboard controls</h2>
+  <table>
+    <tr><th>Control</th><th>Effect</th></tr>
+    <tr><td>Start</td><td>Begins a run: countdown → detect front tab → capture loop.</td></tr>
+    <tr><td>Pause</td><td>Freezes the loop between ticks; leads already saved stay.</td></tr>
+    <tr><td>Resume</td><td>Continues a paused run.</td></tr>
+    <tr><td>Stop</td><td>Ends the run and closes the run record.</td></tr>
+  </table>
+  <p>The event line under the header shows live status; the cards show funnel-stage
+  counts (<code>NEW → ENRICHED → CONTACT_FOUND → EXPORTED</code>).</p>
+
+  <h2 id="trouble">Troubleshooting</h2>
+  <h3>Run stops immediately with "not a supported site / not a profile page"</h3>
+  <p>The front tab wasn't a LinkedIn <code>/in/…</code> profile. Open a profile and Start again.
+  To keep polling instead of stopping, set <code>SCREEN_LEADS_ON_INVALID_PAGE=wait</code>.</p>
+  <h3>"Screen capture failed" or blank/black screenshots</h3>
+  <p>macOS Screen Recording permission isn't granted (or the terminal wasn't restarted after
+  granting). On Linux, you're likely on Wayland — switch to an X11 session.</p>
+  <h3>Page doesn't scroll during capture</h3>
+  <p>macOS Accessibility permission isn't granted, or the profile window isn't focused. Capture
+  still runs; you just get fewer distinct sections.</p>
+  <h3>Authentication / API key errors</h3>
+  <p><code>ANTHROPIC_API_KEY</code> isn't set in the same terminal running <code>python main.py</code>.
+  Re-run the <code>export</code> line, or use <code>ant auth login</code>.</p>
+  <h3>Name captured but no email/phone</h3>
+  <p>Expected — contact details live behind LinkedIn's <strong>Contact info</strong> panel and are
+  often not shown. Open that panel before capture; the lead stays at <code>ENRICHED</code> until
+  contact data is found.</p>
+  <h3>Wrong window captured</h3>
+  <p>Keep the profile tab frontmost. Increase <code>SCREEN_LEADS_START_DELAY</code> to give yourself
+  more time to switch after Start.</p>
+
+  <h2 id="config">Configuration</h2>
+  <p>All optional; set as environment variables (or in a <code>.env</code> file — see
+  <code>.env.example</code>).</p>
+  <table>
+    <tr><th>Variable</th><th>Default</th><th>Purpose</th></tr>
+    <tr><td><code>ANTHROPIC_API_KEY</code></td><td>—</td><td>Required. Claude API key.</td></tr>
+    <tr><td><code>SCREEN_LEADS_MODEL</code></td><td>claude-opus-4-8</td><td>Vision model (use claude-sonnet-5 to cut cost).</td></tr>
+    <tr><td><code>SCREEN_LEADS_START_DELAY</code></td><td>5</td><td>Grace seconds before first capture (0 = off).</td></tr>
+    <tr><td><code>SCREEN_LEADS_SCROLL_STEPS</code></td><td>6</td><td>Screenshots per profile.</td></tr>
+    <tr><td><code>SCREEN_LEADS_SCROLL_AMOUNT</code></td><td>800</td><td>Scroll distance per step.</td></tr>
+    <tr><td><code>SCREEN_LEADS_SCROLL_PAUSE</code></td><td>0.8</td><td>Settle time (s) after each scroll.</td></tr>
+    <tr><td><code>SCREEN_LEADS_WATCH_INTERVAL</code></td><td>2.0</td><td>Loop tick (s).</td></tr>
+    <tr><td><code>SCREEN_LEADS_ON_INVALID_PAGE</code></td><td>stop</td><td><code>stop</code> or <code>wait</code> on a non-target page.</td></tr>
+    <tr><td><code>SCREEN_LEADS_HOST</code> / <code>_PORT</code></td><td>127.0.0.1 / 8000</td><td>Server bind address.</td></tr>
+  </table>
+
+  <h2 id="sites">Adding a site</h2>
+  <ol>
+    <li>Create <code>ai/recipes/&lt;site&gt;.py</code> implementing <code>SiteRecipe</code>
+      (<code>matches</code>, <code>extraction_schema</code>, <code>extraction_prompt</code>,
+      <code>to_canonical</code>).</li>
+    <li>Register it in <code>ai/recipes/__init__.py</code>.</li>
+  </ol>
+  <p>The capture loop and guard are recipe-driven, so nothing else changes.</p>
+
+  <h2 id="compliance">Compliance</h2>
+  <p>Automated <em>scraping</em> of LinkedIn violates its Terms of Service. This tool is built for
+  low-volume, human-in-the-loop use on profiles you manually open and are allowed to view. Keep
+  pacing conservative and use it accordingly.</p>
+</main>
+</body>
+</html>
diff --git a/db/__init__.py b/db/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/db/database.py b/db/database.py
new file mode 100644
index 0000000..323e986
--- /dev/null
+++ b/db/database.py
@@ -0,0 +1,162 @@
+"""SQLite persistence for leads and capture runs.
+
+Thread-safe: the capture loop runs on a background thread while the API serves
+on the main thread, so all access goes through one connection guarded by a lock.
+Leads are keyed by source_url; re-capturing the same profile merges new,
+non-empty fields into the existing row rather than duplicating it.
+"""
+from __future__ import annotations
+
+import json
+import sqlite3
+import threading
+import time
+from typing import Any, Optional
+
+from config import DB_PATH
+from core.models import Lead
+
+_lock = threading.Lock()
+_conn: Optional[sqlite3.Connection] = None
+
+_JSON_FIELDS = ("social_handles", "screenshot_refs", "fields_found")
+
+
+def _connect() -> sqlite3.Connection:
+    global _conn
+    if _conn is None:
+        _conn = sqlite3.connect(str(DB_PATH), check_same_thread=False)
+        _conn.row_factory = sqlite3.Row
+    return _conn
+
+
+def init_db() -> None:
+    with _lock:
+        conn = _connect()
+        conn.executescript(
+            """
+            CREATE TABLE IF NOT EXISTS leads (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                full_name TEXT, first_name TEXT, last_name TEXT,
+                headline TEXT, title TEXT,
+                company TEXT, company_url TEXT, industry TEXT, location TEXT, about TEXT,
+                email TEXT, phone TEXT, website TEXT, social_handles TEXT,
+                source_site TEXT, source_url TEXT UNIQUE,
+                captured_at REAL, updated_at REAL, screenshot_refs TEXT,
+                confidence REAL, fields_found TEXT, needs_review INTEGER,
+                stage TEXT
+            );
+            CREATE TABLE IF NOT EXISTS runs (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                status TEXT, started_at REAL, stopped_at REAL,
+                reason TEXT, leads_captured INTEGER DEFAULT 0
+            );
+            """
+        )
+        conn.commit()
+
+
+def _row_to_lead(row: sqlite3.Row) -> dict[str, Any]:
+    d = dict(row)
+    for f in _JSON_FIELDS:
+        try:
+            d[f] = json.loads(d[f]) if d.get(f) else ([] if f != "social_handles" else {})
+        except Exception:
+            d[f] = [] if f != "social_handles" else {}
+    d["needs_review"] = bool(d.get("needs_review"))
+    return d
+
+
+def get_lead_by_url(source_url: str) -> Optional[dict[str, Any]]:
+    with _lock:
+        cur = _connect().execute("SELECT * FROM leads WHERE source_url = ?", (source_url,))
+        row = cur.fetchone()
+        return _row_to_lead(row) if row else None
+
+
+def upsert_lead(lead: Lead) -> dict[str, Any]:
+    """Insert a new lead, or merge non-empty fields into an existing one."""
+    lead.finalize()
+    with _lock:
+        conn = _connect()
+        existing = conn.execute(
+            "SELECT * FROM leads WHERE source_url = ?", (lead.source_url,)
+        ).fetchone()
+
+        d = lead.to_dict()
+        for f in _JSON_FIELDS:
+            d[f] = json.dumps(d[f], ensure_ascii=False)
+        d["needs_review"] = 1 if d["needs_review"] else 0
+
+        if existing is None:
+            cols = ", ".join(d.keys())
+            ph = ", ".join(["?"] * len(d))
+            conn.execute(f"INSERT INTO leads ({cols}) VALUES ({ph})", tuple(d.values()))
+        else:
+            merged = dict(existing)
+            # prefer new non-empty scalar values; JSON fields already serialised
+            for k, v in d.items():
+                if k in ("id", "captured_at", "source_url"):
+                    continue
+                if k in _JSON_FIELDS:
+                    if v not in ("[]", "{}", "", None):
+                        merged[k] = v
+                elif v not in ("", None, 0, 0.0):
+                    merged[k] = v
+            merged["updated_at"] = time.time()
+            set_clause = ", ".join(f"{k} = ?" for k in merged if k != "id")
+            vals = [merged[k] for k in merged if k != "id"]
+            conn.execute(
+                f"UPDATE leads SET {set_clause} WHERE id = ?", (*vals, merged["id"])
+            )
+        conn.commit()
+    return get_lead_by_url(lead.source_url) or {}
+
+
+def list_leads(stage: Optional[str] = None) -> list[dict[str, Any]]:
+    with _lock:
+        conn = _connect()
+        if stage:
+            cur = conn.execute(
+                "SELECT * FROM leads WHERE stage = ? ORDER BY updated_at DESC", (stage,)
+            )
+        else:
+            cur = conn.execute("SELECT * FROM leads ORDER BY updated_at DESC")
+        return [_row_to_lead(r) for r in cur.fetchall()]
+
+
+def stage_counts() -> dict[str, int]:
+    with _lock:
+        cur = _connect().execute("SELECT stage, COUNT(*) c FROM leads GROUP BY stage")
+        return {r["stage"]: r["c"] for r in cur.fetchall()}
+
+
+def set_stage(lead_id: int, stage: str) -> None:
+    with _lock:
+        conn = _connect()
+        conn.execute(
+            "UPDATE leads SET stage = ?, updated_at = ? WHERE id = ?",
+            (stage, time.time(), lead_id),
+        )
+        conn.commit()
+
+
+# --- runs --------------------------------------------------------------------
+def start_run() -> int:
+    with _lock:
+        conn = _connect()
+        cur = conn.execute(
+            "INSERT INTO runs (status, started_at) VALUES ('running', ?)", (time.time(),)
+        )
+        conn.commit()
+        return int(cur.lastrowid)
+
+
+def finish_run(run_id: int, status: str, reason: str, leads_captured: int) -> None:
+    with _lock:
+        conn = _connect()
+        conn.execute(
+            "UPDATE runs SET status = ?, stopped_at = ?, reason = ?, leads_captured = ? WHERE id = ?",
+            (status, time.time(), reason, leads_captured, run_id),
+        )
+        conn.commit()
diff --git a/main.py b/main.py
new file mode 100644
index 0000000..467f2b6
--- /dev/null
+++ b/main.py
@@ -0,0 +1,18 @@
+"""Entry point: launch the dashboard + API.
+
+    python main.py
+
+Then open http://127.0.0.1:8000 and use Start / Pause / Stop. Requires
+ANTHROPIC_API_KEY in the environment (or an `ant auth login` profile).
+"""
+import uvicorn
+
+import config
+
+
+def main() -> None:
+    uvicorn.run("api.app:app", host=config.HOST, port=config.PORT, reload=False)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..2c50957
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,6 @@
+anthropic>=0.116.0
+fastapi>=0.110
+uvicorn[standard]>=0.27
+mss>=9.0
+pyautogui>=0.9.54
+python-dotenv>=1.0
diff --git a/run.bat b/run.bat
new file mode 100644
index 0000000..071c407
--- /dev/null
+++ b/run.bat
@@ -0,0 +1,28 @@
+@echo off
+REM One-command runner for Screen Leads (Windows).
+cd /d "%~dp0"
+
+if not exist ".venv" (
+  echo Creating virtual environment (.venv)...
+  python -m venv .venv
+)
+call .venv\Scripts\activate
+
+if not exist ".venv\.installed" (
+  echo Installing dependencies...
+  pip install -q --upgrade pip
+  pip install -q -r requirements.txt
+  type nul > .venv\.installed
+)
+
+if "%ANTHROPIC_API_KEY%"=="" if not exist ".env" (
+  echo ANTHROPIC_API_KEY is not set and no .env file was found.
+  echo   set ANTHROPIC_API_KEY=sk-ant-...   ^(this terminal^)
+  echo   or copy .env.example to .env and edit it
+  exit /b 1
+)
+
+if "%SCREEN_LEADS_HOST%"=="" set SCREEN_LEADS_HOST=127.0.0.1
+if "%SCREEN_LEADS_PORT%"=="" set SCREEN_LEADS_PORT=8000
+echo Starting Screen Leads at http://%SCREEN_LEADS_HOST%:%SCREEN_LEADS_PORT%  (Ctrl-C to stop)
+python main.py
diff --git a/run.sh b/run.sh
new file mode 100755
index 0000000..8b6a26c
--- /dev/null
+++ b/run.sh
@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# One-command runner for Screen Leads (macOS / Linux).
+# Creates the virtualenv, installs deps on first run (or when requirements
+# change), checks for the API key, and launches the dashboard.
+set -euo pipefail
+cd "$(dirname "$0")"
+
+PYTHON="${PYTHON:-python3}"
+
+# 1. Virtual environment
+if [ ! -d ".venv" ]; then
+  echo "→ Creating virtual environment (.venv)…"
+  "$PYTHON" -m venv .venv
+fi
+# shellcheck disable=SC1091
+source .venv/bin/activate
+
+# 2. Dependencies — (re)install only when requirements.txt is newer than the marker
+if [ ! -f ".venv/.installed" ] || [ requirements.txt -nt ".venv/.installed" ]; then
+  echo "→ Installing dependencies…"
+  pip install -q --upgrade pip
+  pip install -q -r requirements.txt
+  touch ".venv/.installed"
+fi
+
+# 3. API key — config.py also loads a .env file, so either is fine
+if [ -z "${ANTHROPIC_API_KEY:-}" ] && [ ! -f ".env" ]; then
+  echo "⚠  ANTHROPIC_API_KEY is not set and no .env file was found."
+  echo "   Fix with one of:"
+  echo "     export ANTHROPIC_API_KEY=sk-ant-...     (this terminal)"
+  echo "     cp .env.example .env && edit it          (persisted)"
+  exit 1
+fi
+
+# 4. Launch
+HOST="${SCREEN_LEADS_HOST:-127.0.0.1}"
+PORT="${SCREEN_LEADS_PORT:-8000}"
+echo "→ Starting Screen Leads at http://${HOST}:${PORT}  (Ctrl-C to stop)"
+exec python main.py
diff --git a/setup.bat b/setup.bat
new file mode 100644
index 0000000..7201768
--- /dev/null
+++ b/setup.bat
@@ -0,0 +1,41 @@
+@echo off
+REM One-time setup for Screen Leads (Windows):
+REM   1. create the virtual environment
+REM   2. install dependencies
+REM   3. prompt for environment variables and write .env
+setlocal enabledelayedexpansion
+cd /d "%~dp0"
+echo == Screen Leads setup ==
+
+if not exist ".venv" (
+  echo Creating virtual environment (.venv)...
+  python -m venv .venv
+) else (
+  echo Virtual environment already exists.
+)
+call .venv\Scripts\activate
+
+echo Installing dependencies...
+pip install -q --upgrade pip
+pip install -q -r requirements.txt
+type nul > .venv\.installed
+
+if exist ".env" (
+  echo .env already exists - leaving it untouched.
+  goto done
+)
+
+echo.
+echo Set up your environment variables (saved to .env):
+set /p API_KEY="  Anthropic API key (sk-ant-...): "
+set /p MODEL="  Model [claude-opus-4-8] (Enter to accept): "
+if "!MODEL!"=="" set MODEL=claude-opus-4-8
+> .env echo ANTHROPIC_API_KEY=!API_KEY!
+>> .env echo SCREEN_LEADS_MODEL=!MODEL!
+echo Wrote .env
+if "!API_KEY!"=="" echo   WARNING: no key entered - edit .env before running.
+
+:done
+echo.
+echo Setup complete. Start the app with: run.bat
+endlocal
diff --git a/setup.sh b/setup.sh
new file mode 100755
index 0000000..c69849d
--- /dev/null
+++ b/setup.sh
@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# One-time setup for Screen Leads (macOS / Linux):
+#   1. create the virtual environment
+#   2. install dependencies
+#   3. prompt for environment variables and write .env
+# After this, start the app with ./run.sh
+set -euo pipefail
+cd "$(dirname "$0")"
+
+PYTHON="${PYTHON:-python3}"
+echo "== Screen Leads setup =="
+
+# 1. Virtual environment
+if [ ! -d ".venv" ]; then
+  echo "→ Creating virtual environment (.venv)…"
+  "$PYTHON" -m venv .venv
+else
+  echo "→ Virtual environment already exists."
+fi
+# shellcheck disable=SC1091
+source .venv/bin/activate
+
+# 2. Dependencies
+echo "→ Installing dependencies…"
+pip install -q --upgrade pip
+pip install -q -r requirements.txt
+touch ".venv/.installed"
+
+# 3. Environment variables → .env
+if [ -f ".env" ]; then
+  echo "→ .env already exists — leaving it untouched."
+else
+  echo
+  echo "Set up your environment variables (saved to .env):"
+  printf "  Anthropic API key (sk-ant-…): "
+  read -rs API_KEY; echo
+  printf "  Model [claude-opus-4-8] (Enter to accept): "
+  read -r MODEL
+  MODEL="${MODEL:-claude-opus-4-8}"
+  {
+    echo "ANTHROPIC_API_KEY=${API_KEY}"
+    echo "SCREEN_LEADS_MODEL=${MODEL}"
+  } > .env
+  chmod 600 .env
+  echo "→ Wrote .env"
+  if [ -z "${API_KEY}" ]; then
+    echo "  ⚠  No key entered — edit .env and set ANTHROPIC_API_KEY before running."
+  fi
+fi
+
+echo
+echo "✓ Setup complete. Start the app with:  ./run.sh"