Pomerado Turns Websites Into Agent APIs, Claims 7x Speedups

A Y Combinator-backed startup called Pomerado went live on Hacker News with a pitch that its service can convert almost any website into a machine-readable API for AI agents without waiting for the site owner to publish one. The company says that approach is both quicker and steadier than letting an agent click through pages in a browser, and it published benchmark figures to back that up.

Why it matters: most useful web software still has no public API, so agents that need to read a fare table, an invoice list or an insurance claim status have to either drive a browser or give up. If a hosted layer can turn those pages into callable tools reliably, the practical bottleneck for agent work shifts from "is there an API?" to "who is allowed to call it, and with whose login?"

What did Pomerado actually launch?

According to the company's launch post, Pomerado exposes an MCP server that coding agents can register in one command. For Claude Code the setup is a claude mcp add --transport http pomerado call against a keyless endpoint; for Codex it is a codex mcp add pomerado --url call. The company says no account is needed for a trial, and suggests starting with a read-only task such as finding a tool for "search flights" and running it.

The underlying model is not a new protocol but a hosting arrangement. Pomerado describes itself as open source, says it builds a deterministic API for a target site in real time, and states that every integration arrives hosted, scaled and access controlled, so a team does not stand up its own infrastructure. Its dashboard lists integrations as collections of named tools, each tagged with whether it needs a saved login.

How does it compare with browser agents, according to Pomerado?

The company's central claim is speed and cost. It says customers have observed roughly 7x speedups, on the argument that browser agents reason their way through every page while Pomerado's generated APIs run in seconds.

The example it leads with is an employee-onboarding task spanning HR, IT, course and library systems. Pomerado says it finished that in 8 minutes 16 seconds at a cost of 3 cents, while Codex Computer Use took 21 minutes 22 seconds and $41.01. The post notes the demo video is shown at 321x speed.

On benchmarks, Pomerado says it was 2x faster, 10 to 1,000 times cheaper and 30% more accurate than Astra Computer Use, Instinct and Muse for browser use. That comparison covered 30 tasks across 25 business applications, one attempt each, over a list that includes OpenEMR, ERPNext, Moodle, Koha, SuiteCRM, QuickBooks Online, osTicket and Paperless-ngx. The company is explicit about one asymmetry: its cost figure counts model tokens at list price, and it says Instinct and Muse do not report theirs. All of these numbers come from Pomerado itself and have not been independently verified.

What happens when a website changes?

This is the part that separates Pomerado from a one-off scraping script. The post describes a "self-healing" loop: when a page changes, the integration fails, Pomerado detects the break, runs a maintenance pass, patches the failing step, tests it and republishes, and the original run is then resumed.

Its worked example is an Anthem claims lookup. The date-of-service filter failed because Anthem had moved it into a new Search options panel. Maintenance repaired the filter in about four minutes, the test returned 24 claims, and the original run was updated with 19 paid, 3 pending and 2 denied for missing prior authorization. Whether that recovery holds up across many sites and many months is an open question the launch post does not settle.

Can an agent build the integration itself?

Yes, that is the second workflow Pomerado shows. Rather than writing the integration by hand, you ask a coding agent such as Codex, Claude or Cursor to create one, and Pomerado builds and tests it while the agent answers its clarifying questions.

In the Coupa supplier-portal example, the agent calls a tool-existence check, starts a build, and Pomerado asks back about scope: whether to include all three customers invoiced on that account, and whether to pull comment threads on disputed invoices. The agent answers, the build is tested against a date range, and the resulting read-only tool is published and run once. The company says this works across prior authorization portals, EMRs, ERPs, CRMs, payer and claims systems, government filings, procurement and bank portals.

How does Pomerado handle logins and 2FA?

Authentication is treated as part of the service rather than the customer's problem. Pomerado says it manages sign-in for each integration, including sites that require two-factor codes or one-time passwords, and its Logins screen shows masked usernames with a verified status and the 2FA method per site, covering SMS, authenticator app and email codes, plus saved recovery codes.

On the FAQ, the company says you do not have to store credentials at all: logins can be passed with each API request. It also says it retains hashes of sign-in identifiers to enforce its one-account-per-website limit, and that it enforces rate limits and handles anti-bot measures, including residential proxies. Those are the claims a security reviewer should press on before wiring this into anything that can take actions rather than just read.

Where does this fit in Grafy?

The closest analogue in Grafy is Operator, the browsing agent that performs web tasks on your behalf. Pomerado's argument is essentially a critique of that category: instead of an agent improvising through a UI, precompute a stable tool for the site and call it. If you want to test that trade-off yourself, Operator is the place to feel the cost of open-ended browsing, and Grafy Chat is a reasonable place to compare how different models plan the same web task before you commit an agent to it.

For the research angle, Reader can work through the benchmark methodology in Pomerado's docs and press release and tell you which comparisons are like-for-like and which rest on numbers only one vendor reported. And if you are building the integration layer yourself, Grafy Code is where you would write and run the client code against an MCP endpoint like the one Pomerado publishes.

Frequently asked questions

Does Pomerado need an API from the website it targets?

No. That is the premise of the launch: Pomerado says it builds an API or MCP interface for a site that has none, covering legacy portals, consumer sites and internal tools. It generates a deterministic integration for the target pages, hosts it, and maintains it when the site changes. The trade-off is that you are automating access to a site on terms the site owner did not publish, which is a question for your own legal and security review.

Is Pomerado faster and cheaper than browser agents?

That is Pomerado's claim, not an independent finding. The company reports about 7x speedups, 2x more speed and 10 to 1,000 times lower cost than Astra Computer Use, Instinct and Muse on 30 tasks across 25 business apps, and its onboarding demo ran in 8 minutes 16 seconds for 3 cents against 21 minutes 22 seconds and $41.01 for Codex Computer Use. It also notes competitors do not report comparable cost figures.

What does Pomerado cost to use?

The post lists three tiers. Free is $0 per month with 100 tool calls per day and 10 new integrations per month. Premium is $50 per month, monthly or yearly, with 1,000 tool calls per day and 100 new integrations per month. Business is quoted on contact and adds higher call and integration volumes, unlimited accounts per website, custom 2FA and proxying, MCP code ownership, private integrations, roles, audit logs, SLAs and 24/7 support.

---

Source: Hacker News — front page — Launch HN: Pomerado (YC F26) – Turn any website into an MCP for agents