Pomerado, a Y Combinator Fall 2026 Startup, Sells Agents Access to Sites With No API for $50 a Month

Pomerado, a two-person Y Combinator Fall 2026 company, charges $50 a month for 1,000 tool calls a day to let AI agents operate websites that have no API. Neither its site nor its Y Combinator listing says how much it has raised, so the only price anyone can check is the one on the pricing page.

What Pomerado sells and who runs it

The company launched on Hacker News with a Launch HN post that reached 63 points. Its Y Combinator listing names Akshay Manglik as founder and chief executive, previously an Applied AI Engineer at Scale AI, and Alan Ma as founder, previously a Business Analyst at McKinsey. The team is two people in San Francisco, and the primary partner is Aaron Epstein.

The product records an agent completing a task on a site once, then turns the run into a repeatable tool exposed as an MCP server, the protocol coding agents use to call outside tools. Later calls run that stored tool instead of a model, which the founders say usually returns in 5 to 30 seconds. A failed run triggers a repair agent.

Pricing: free, $50 and "contact sales"

Tier	Price	Tool calls a day	New integrations a month
Free	$0	100	10
Premium	$50 a month	1,000	100
Business	Contact sales	100,000 or more	1,000 or more

Every tier is limited to one account per website except Business, which is unlimited. Business adds audit logs, service-level agreements and private team integrations, according to the pricing page.

The benchmark is the company's own

The page says Pomerado is 2 times faster, 10 to 1,000 times cheaper and 30% more accurate than Astra Computer Use, Instinct and Muse. The test covers 30 tasks across 25 business apps with one try each, and Pomerado supplied it. No independent party is named as having run it.

1. Pomerado496 s
2. Codex Computer Use1282 s

Source: Pomerado pricing page, accessed 2026-11-14

The page's worked example is where the claims strain. Pomerado finished the onboarding task in 8 minutes 16 seconds for 3 cents. Codex Computer Use took 21 minutes 22 seconds and cost $41.01. That is a cost ratio of about 1,367, above the 1,000 ceiling the page gives for the same comparison.

Licence, requirements and what the README admits

The GitHub repository carries 1.7k stars, 146 forks, 22 open issues and 51 open pull requests. It is MIT licensed, though versions 0.1.2 and earlier were AGPL-3.0-only. Building an integration needs an OpenAI API key, Node 24.21 or a later Node 24 release and a Chromium install through Playwright. Running a finished integration makes no model request, the README says.

The README is candid about limits. Pomerado does not check where a page's own scripts send what a user types. Write-authority integrations can change data on a site, and a run does not check authority or intent. The README says nothing about the terms of service of the sites an integration drives.

Competitors and the questions from the thread

The founders contrast the product with WebMCP, which needs a site's cooperation, and with Printing Press, which generates command-line tools. A commenter asked why OpenAI or Anthropic would not build this natively. The founders answered that the anti-bot infrastructure and token savings make it a different approach.

A team member, writing as am03, said Amazon works through custom stealth tooling, in-house CAPTCHA solvers and cloud browsers. The founders reported a 2.2% run error rate, with 82% of errors resolved automatically, and said failures cluster on travel and automotive marketplaces. Both figures are founder claims in a comment thread.

Other agent tooling takes different routes, such as Brave's API that takes USDC per request and Chrome's Modern Web Guidance skills for coding agents.

The repository's 22 open issues and 51 open pull requests are where the repair-agent claims will be tested first. The question the README leaves blank is whether a site that bans automated access can enforce its terms against a customer's agent.