The problem surfaces almost immediately. An AI agent navigates to a website, parses the HTML, identifies what looks like a checkout button—and then the session times out. Or two-factor authentication kicks in. Or the site redesigns overnight and every CSS selector breaks. What should be straightforward automation becomes, in practice, a game of probabilistic whack-a-mole that eats tokens and burns engineering hours.
Rindler, a startup out of Y Combinator's Summer 2026 batch, thinks it has found a way around the chaos: stop making agents parse websites at all. Instead, turn those websites into something agents can understand—stable, deterministic APIs.
It's an ambitious premise. And whether it works may say something larger about whether agentic workflows can ever escape the demo stage.
The Core Bet: Mapping Over Parsing
Michael Serrano and Arthur De Los Santos, both MIT graduates who founded the Boston-based company in 2025, describe Rindler as "the translation layer between AI agents and the web." The pitch is technical but the frustration it addresses is universal among developers building with agents: every time you scrape a site, you're re-parsing DOM structures, chasing selectors, hoping the layout hasn't shifted since your last run.
Rindler's approach is different, at least in theory. The platform maps a website once—breaking it down into structured screens, records, and typed actions. After that initial mapping, agents don't touch raw HTML. They call stable API endpoints: search, add_to_cart, download_records. The company describes these as "self-healing configs" that adapt when websites change, though exactly how that self-healing works in practice isn't spelled out in public documentation.
Authentication, often the breaking point for web automation, runs through a real server-side browser. Users log in and complete two-factor authentication inside a managed session; Rindler encrypts and stores session cookies using AES-256-GCM encryption with per-user keys. Once authenticated, agents interact through a hosted Model Context Protocol server at mcp.rindler.ai, compatible with Claude Code, Cursor, ChatGPT, and other MCP-enabled frameworks.
The performance claims are bold. Rindler says the system delivers 3x fewer failed tasks, 4x faster workflow completion, and 6x lower execution costs compared to traditional scraping or scripted browser automation. Those numbers come from the company itself; no independent benchmarks exist yet.
What's Live, What's Not
As of a recent snapshot in early August 2026, Rindler's demo catalog had 19 pre-mapped sites. The roster spans banking (Chase, Bank of America), travel (United Airlines, Airbnb), e-commerce (Amazon, Instacart, DoorDash), professional networks (LinkedIn), and state business registries. Some of these workflows are running in production. Others remain mapped but dormant.
The company has built multi-tenant applicant tracking system integrations for Greenhouse, Lever, and Ashby—tools used by recruiting teams to manage candidates. State business registry lookups are live. Banking flows exist but require manual verification on a per-customer basis; Chase statement downloads work end-to-end, though setup isn't automated.
Then there's the harder stuff. Court e-filing systems, PACER access for federal case records, supplier and accounts payable portals, healthcare payer systems covering eligibility checks and prior authorization—all of these appear in Rindler's technical documentation, last updated August 6, as mapped but not yet in automated production. Healthcare and insurance workflows are "available on request," the docs note, but aren't running live. Rindler has not obtained SOC 2, HIPAA, or PCI certification, which would be table stakes for many enterprise buyers in those sectors.
It's the classic startup tension: showing enough capability to attract customers without overpromising on infrastructure that isn't fully baked.
Pricing, Practicalities, and Legal Gray Zones
Rindler's pricing is straightforward. The Starter plan runs $100 per month with a seven-day trial, covering one login-gated site plus access to the pre-mapped catalog and 100 one-off runs monthly. Teams pricing jumps to $1,000. Enterprise is custom. No free tier.
Installation is initiated using curl https://rindler.ai/install | sh. No API key required; authentication flows through the hosted MCP server.
The legal positioning is careful but raises questions. Rindler acts only on sites the customer already has access to, using the customer's own credentials. The company places the burden of compliance with third-party terms of service squarely on the customer. If a vendor objects—and some inevitably will—Rindler says it will stop supporting that site immediately. It's a reactive stance, not a proactive one. Whether that holds up as the startup scales remains to be seen.
The Landscape They're Entering
Rindler's launch comes at a moment when the Model Context Protocol, now governed by the Linux Foundation's Agentic AI Foundation with backing from Anthropic, Block, and OpenAI, is gaining real traction. MCP aims to standardize how agents interact with external systems—think of it as the connective tissue for agentic workflows.
Several other startups are circling the same territory. Unbrowse, Anysite.io, and YC alum Wildcard are all building in the "website-to-API" or agent-infrastructure space, each with slightly different technical bets. The market is early enough that no single approach has proven itself yet.
Serrano, the CEO, brings machine learning engineering experience from Roblox and LLM research work at MIT's Computer Science and Artificial Intelligence Laboratory. De Los Santos, who graduated from MIT in 2026 with a focus on CS and AI, rounds out the founding team. Y Combinator's directory lists the team size as two, though LinkedIn indicates a range of 2–10 employees and approximately 306 followers in July 2026. No external funding rounds have been publicly announced beyond YC as of August 9, 2026.
The Bigger Question
Strip away the technical architecture and what Rindler is really asking is this: can web automation for AI agents move from probabilistic to deterministic? Can workflows that currently break half the time become reliable enough for operations teams to actually trust them?
The answer isn't just about Rindler. It's about whether the current wave of agentic AI can graduate from demos—where failure is tolerable, even charming—to production systems where downtime has consequences. The startup has built an interesting technical layer. Whether it solves the underlying problem, or just shifts it, is something only real-world usage will reveal.
For now, the founders are mapping websites, one login flow at a time, betting that determinism beats guesswork. In the messy world of web automation, that's not a bad bet to make.
