What changed
FACT: Google shipped computer use in Gemini 3.5 Flash, a low-cost/low-latency tier (deepmind.google blog). FACT: 'Fortress' launched an engine-level fingerprint-spoofed Chromium exposed over MCP that reportedly defeats top bot detectors and is consumable directly by agents (Show HN / tilion.dev). HYPOTHESIS: together these drop the per-task cost of reliably browsing anti-bot sites low enough that solo builders can run always-on monitors economically.
Why now
HYPOTHESIS: prior to a cheap Flash-tier computer-use model, autonomous browsing of JS-heavy, bot-hostile sites was either expensive (frontier model per page) or brittle (headless gets blocked). The stealth-Chromium-over-MCP + cheap-model pairing is what makes per-task monitoring cross the cost/reliability line. This is a capability window, not a demand event.
Converging signals
Two independent capability signals — (1) cheap agentic computer use (Gemini 3.5 Flash), (2) MCP-native undetectable browser (Fortress) — converge on the same unlock: economically viable autonomous access to sites that actively block bots. No demand signal converges with them here.
Customer pain
HYPOTHESIS ONLY — no demand_evidence was provided. Plausible pains (unverified): buyers who need to know the moment a permit is issued, a court record is filed, a competitor changes price, or a scarce SKU restocks, and who currently can't get it because the source site blocks scrapers. None of this is evidenced in the input; it must be validated before build.
Who pays
HYPOTHESIS: a narrow vertical buyer, e.g. contractors/expediters tracking building-permit and inspection portals; litigation-support or skip-trace shops tracking court dockets; resellers tracking marketplace stock/price. The generic 'anyone who scrapes' buyer is NOT a real buyer and should be ignored.
Solved today
FACT-adjacent: incumbents exist — commercial scraping/proxy platforms (Bright Data, Apify, Zyte), price/stock monitors (Visualping, Distill, keepa-style trackers), and DIY headless scripts. HYPOTHESIS: these are either priced for volume/enterprise or get blocked on the hardest sites.
Why current solutions are bad
HYPOTHESIS: general scraping platforms don't package a vertical (they hand you raw HTML, not 'a new permit was issued at 123 Main St'); cheap DIY monitors get blocked by modern bot detection; the value is in the vertical interpretation + alerting, which nobody sells cheaply for these specific niches.
Proposed product
A single-vertical monitoring micro-SaaS: pick ONE bot-hostile source and one paying niche (recommend building-permit / inspection portals given founder's public-records + industrial-ops edge), run the stealth-browser+cheap-model stack to poll it, and sell structured change-alerts (email/SMS/webhook) on a per-seat or per-watched-entity subscription. Sell the outcome (the alert), keep the scraping tech invisible.
MVP version
One source portal + one detector: agent logs in / navigates the target site via stealth Chromium (MCP), extracts the target records with the cheap model, diffs against last run, and fires a structured alert. Manual customer onboarding. Postgres for state (you already run one). No self-serve signup needed for first 5 customers.
30-day build
Validate demand FIRST (this is the gating risk): interview 15-20 candidates in the chosen niche, confirm a specific unmet pain and willingness to pay, and identify the exact source portal(s). In parallel, spike the stealth-browser+MCP+Flash stack against that one portal to confirm it isn't blocked and cost-per-poll is acceptable. Kill or pivot if either fails.
60-day build
Build the single-vertical MVP for 3-5 design-partner customers found in the interviews; instrument reliability (block rate, false-positive alerts) and cost per watched entity. Charge from day one even if manually operated.
90-day revenue plan
Convert design partners to paid ($X/entity/month or $/seat), add self-serve for that one vertical, and only THEN consider a second source/vertical. Target first revenue from the manual/design-partner phase inside 60-90 days, contingent on demand validating.
Distribution path
HYPOTHESIS: direct outbound to the specific niche (trade groups, subreddits, industry Slack/FB groups, cold email to named firms), demonstrated-value demos (show the alert firing on their real target), not ads. Founder sells through demonstrated value, which fits.
Pricing hypothesis
HYPOTHESIS: usage-aligned — $29-99/mo per seat for a light watcher, or per-watched-entity (e.g. $X per address/docket/SKU) for heavier users. Price on value of timeliness, not on compute cost.
Technical difficulty
Moderate. The stealth-browser+MCP+model stack is largely off-the-shelf; the hard parts are reliability against a site that will change its anti-bot posture, keeping cost per poll low, and low false-positive alerting. Solo-buildable.
Legal / regulatory risk
ELEVATED and the central non-demand risk. Deliberate anti-bot/fingerprint evasion can violate target-site ToS and, depending on the site and data, implicate CFAA-style access claims and IP/scraping case law. Public-government portals (permits/court records) are the safer end (public records, weaker ToS claims); logged-in marketplaces are the riskier end. Choose public-data verticals to de-risk. Not legal advice.
Platform dependency
HIGH: the product rides on (a) the third-party stealth browser (Fortress/tilion — early, single-vendor, may change terms or fold), (b) a specific model tier, and (c) the target site's tolerance. Any of the three can break the product. Mitigate by abstracting the browser layer and choosing durable public sources.
Founder fit
Moderate-to-good but NOT his best shape. The public-records/permit-monitoring vertical aligns strongly with his public-records, industrial-ops, and compliance-monitoring strengths and his demonstrated-value sales style. However, this is a scraping/ToS-risk play, not the government-portal FILING-mandate shape (forced buyer, per-transaction fee) where he has proven, higher-fit success. If a permit-monitoring wedge can be paired with a filing/compliance obligation, fit rises sharply.
Breakout potential
Moderate. If one vertical monitor lands, the same engine generalizes to many niches (court records, licensing boards, procurement postings, marketplace stock). But the underlying stealth-browser layer is commoditizing fast, so defensibility lives in the vertical data model + distribution, not the tech.
Final recommendation
CONDITIONAL / NARROW — do not build the generic 'monitor any site' product. Treat this as a lead to ONE public-data vertical (recommended: building-permit / inspection or court-record monitoring), and gate all build spend behind a 30-day demand-validation sprint. If real buyers with willingness-to-pay surface AND a filing/compliance angle can be attached, escalate; if demand doesn't validate, shelve as 'revisit later.' As presented (no demand evidence, high platform/legal risk, commodity core), it is NOT a strong buy.
Next action
Run a 2-week demand probe on the strongest single vertical (building-permit monitoring): interview 15 target buyers (permit expediters, contractors, GIS/records firms) to confirm a specific unmet pain and a price they'll pay, and simultaneously spike the stealth-browser+MCP+Flash stack against one real permit portal to confirm it isn't blocked and cost-per-poll is viable.