Shell jacket
$189
Search products…
Cart (2)Winter collection
Layered kits for cold commutes
Merino base
$64
Trail pack
$112
Insulated cap
$38
6.5x
Token savings in published search benchmark
35
Sites tested in the public coverage matrix
MCP
Ready for Claude Desktop and Cursor tool loops
northline.store/products
6.5×
Less context sent
Measured menu vs vision screenshot on DuckDuckGo search workflow
77%
Sites successfully indexed
27 of 35 sites in our coverage matrix - not universal
459
Tokens required for navigation
Search workflow menu payload - vs ~3,000 for a screenshot step
Before vs after
Same page. One path sends pixels and hopes. The other sends a structured action space agents can execute deterministically.
Traditional browser agent
- 01Screenshot
- 02Analyze
- 03Guess
- 04Click
- 05Retry
- Context
- ~3000 tokens / step
- Latency
- Vision + retry loops
- Determinism
- Coordinates & pixels
- Reliability
- Opaque failures
Sentinel agent
- 01Open
- 02Observe
- 03Act
- 04Done
- Context
- 459 tokens (measured)
- Latency
- 12.6s end-to-end search workflow
- Determinism
- click 2 · type 1 · submit
- Reliability
- 83% action pass (35 sites)
What Sentinel is
A semantic interaction layer between LLM agents and the browser. We convert pages into structured, executable menus - the agent receives text like [2] click · Sign in → accounts.site.com instead of interpreting screenshots.
Shipped today as CLI, MCP server (Claude Desktop / Cursor), and a desktop observability app. This is early - we publish real test numbers, not marketing fiction.
The problem we solve
Vision loops are expensive
Screenshot context for a single step often costs 1,500-3,000 tokens. Long agent sessions burn budget on pixels, not decisions.
Selectors break silently
CSS paths and coordinate clicks fail when layouts shift. Agents retry blindly with no structured error to act on.
Not every page is a form
Dynamic result grids, CAPTCHAs, and visual-only UIs need different tools. Sentinel says so instead of pretending.
How Sentinel addresses it
- 01Scan the DOM → classify controls by role → assign stable indices
- 02Agent acts with click / type / submit by index - full index stays in memory
- 03Smart menu cap (~720 tokens) + browser_find for hidden controls
- 04Classified failures (CAPTCHA, overlay, not indexable) with fallback guidance
Where Sentinel sits
Agents never touch raw DOM. They talk to Sentinel over MCP; Sentinel owns the browser and returns numbered action menus.
Agent clients
- Claude Desktop
- Cursor
- Claude Code
- Custom MCP host
Sentinel
- MCP server · stdio
- Scan · index · act
- Smart menu + browser_find
Browser layer
- Playwright + Chrome
- Persistent profile
- Live DOM access
Website
- Any public page
- Forms · search · nav
- CAPTCHA & overlays = classified failures
Install once via MCP config. No hosted browser farm - runs locally on your machine with your profile and credentials.
Measured results
Every figure below comes from automated test runs on real sites. We do not claim 95% reliability or 50× savings globally.
Token cost · search workflow
DuckDuckGo · Search query → submit → inspect results menu
Sentinel menu (measured)459
Vision screenshot estimate (high detail)3,000
6.5× cheaper · 84.7% reduction · measured 2026-06-18
35-site coverage matrix
Last run 2026-06-18
Scan pass (27/35)77%
Action probe pass (29/35)83%
Failing sites10 / 35
Mean latency per site: 10,849ms · Multi-agent stress test skipped (budget)
Pass rate by category
Search (5/5)100%
Education (2/2)100%
SaaS (5/6)83%
Developer (4/5)80%
News (4/5)80%
Travel (2/3)67%
E-commerce (3/6)50%
Community (1/2)50%
Content (0/1)0%
Smart menu cap
Curated agent payload vs full index vs ~1,500 token screenshot baseline
Smart
248
Full
248
Vision~
1,500
GitHub
Smart
366
Full
1,221
Vision~
1,500
Smart
549
Full
2,945
Vision~
1,500
Failure patterns
No failure25
Bot / CAPTCHA5
Low scan confidence2
Overlay blocked2
Network error1
What it can do today
Sentinel is a control plane for structured browser interaction - not a general replacement for vision, web search, or travel APIs.
Semantic scan & index
Discover buttons, links, inputs, and modern interactive patterns. Assign data-sfa-index handles that survive rescan.
Agent menu payload
Compact lines: [n] action · label → destination. Smart cap keeps dense pages under ~720 tokens while full index stays in memory.
Deterministic actions
click, type, hover, submit by index. Multi-step sequences with timeouts, retries, and stale-menu guards.
browser_find
Search the full indexed page when a control is hidden from the curated menu - not just the visible subset.
browser_read
Extract main article/body text after navigation - menus list controls, not page content.
Graceful degradation
Classified failures: CAPTCHA, auth required, overlay blocked, site not indexable - with actionable fallback hints.
Strong fit
- +Web search → open result → read content
- +Login, signup, and multi-step forms
- +SaaS dashboards and developer docs
- +Repeatable agent workflows with audit trails
- +MCP / Claude Desktop complex tool chains
Weak fit today
- −Flight or hotel price shopping (dynamic result grids)
- −CAPTCHA-heavy or bot-blocked sites
- −Visual-only UIs (maps, canvas, seat pickers)
- −“Compare every price on the page” without APIs
Pipeline
After MCP connects, every page interaction follows the same loop. See architecture for where Sentinel sits in your stack.
- 01
Scan
Playwright loads the page. Sentinel discovers interactive controls and tags them with data-sfa-index.
- 02
Index
Elements classified by role (search_input, button, link). Smart cap sends top ~42 items; full index kept for find + act.
- 03
Act
Agent calls browser_act({ index, action, value }). Executor clicks/types with retries and stale-menu detection.
- 04
Re-observe
After navigation or DOM change, fresh menu returned. Mutation version blocks actions on remapped indices.
Config
MCP tool surface
# Claude Desktop - claude_desktop_config.json{ "mcpServers": { "sentinel-sfa": { "command": "node", "args": ["/path/to/sentinel-sfa/src/mcp/server.js"] } }}# Tools# browser_open · browser_observe · browser_scroll# browser_find · browser_act · browser_act_sequence# browser_read · browser_closeSite coverage
These destinations passed scan and safe action probes in our latest matrix run. Ten sites failed - CAPTCHA, overlays, low confidence, or network. We publish both lists.
GoogleBingDuckDuckGoWikipediaYahooGitHubGitLabMDNnpmCourseraKhan AcademyNotionSlackFigmaDropboxTrelloBBCCNNNY TimesAmazoneBayTargetAirbnbBookingReddit
Known failures
- ·Medium, Product Hunt, Stack Overflow - bot / CAPTCHA
- ·Best Buy - bot block
- ·Etsy - low scan confidence
- ·Walmart, The Guardian - consent overlay blocked action
- ·Reuters - low scan confidence
- ·Canva - bot block
- ·Tripadvisor - network timeout
Browser is the demo. The model is the product.
Sentinel is not a flight-search app or a Chrome plugin. It is infrastructure for agents to interact with structured surfaces - starting with the web because it is the most accessible testbed.
Now
Browser semantic layer
CLI, MCP, desktop app. Prove indexed control works on real sites with honest metrics and degradation paths.
Next
Deeper page understanding
Accessibility-tree indexing, SPA settle, structured region read, overlay dismissal - without defaulting to vision.
Long term
OS-wide agent interaction
Same abstraction model beyond the browser: numbered semantic menus for native apps, not screenshots at OS scale.
Pre-production: not claiming production-ready universal coverage. Early access targets teams building agent products who need a deterministic browser control plane today.
FAQ
A semantic browser control layer for AI agents. It scans pages, builds numbered menus of interactive controls, and executes click/type/submit by index via MCP, CLI, or desktop app. It is pre-production infrastructure, not a consumer travel or shopping bot.