Image: GitHub
Why it mattersAn agent that calls a saved request runs faster and reads fewer tokens than one that drives Chrome, so a team that already pays an agent to use a website every day can cut the bill without changing the task.
Any website an agent reaches today is a browser the agent has to drive, token by token, every time.
api-anything, released under MIT by GitHub user goodnight000 on 27 September 2026, records the front-end request a site's own page makes once, then lets an agent call that work over plain HTTP from then on. The repository had 420 stars at the time of writing, 13 days after it went up. It ships with 22 read operations across nine sites, including Google Flights, Hacker News, Goodreads and LinkedIn.
What the measurement actually says
The project's README reports two benchmark runs it published itself, so these are the author's own numbers rather than independent ones. On Google Flights, over five runs, the agent calling api-anything answered a one-date search in a median of 8.4 seconds against 12.9 for the same agent driving Chrome through Playwright MCP, at $0.074 against $0.166 for the same request. For five dates, the gap widened: 21.1 seconds against 41.6, at $0.152 against $0.154. All 20 answers were correct across both tools. On Goodreads, where the public API has been gone since 2020, an agent returned five books in 30.5 seconds against 182.5 for the browser agent, and one book cost about a third as much.
Teaching the Goodreads site took 6.3 minutes and $1.19 of model cost in a single run, which produced four operations. The test ran against 30 books the exploring agent had never seen. For the 29 books search found, every title, rating, rating count and page count matched the page Chrome rendered, across 146 calls, all over plain HTTP with no browser involved.
How the teaching step works
The user runs api-anything add <site> with the page URL, a placeholder for the input, and two examples. Headless Chrome loads the page once per example. The part of the request that changes with the input becomes a parameter; the rest is stored as a template. No model is used in building the template, which is the piece that makes a saved operation cheap to call later. Credentials for sites that need an account are imported from a Chrome, Arc, Brave, Edge, Chromium or Firefox profile the user is already signed in to, and saved operations carry references rather than values.
Through the Model Context Protocol, an agent calls operations with four tools: list_sites, list_operations, call_operation and login. The call falls through three layers in order: Node's fetch against the saved request, the same request from inside a Chrome page if the site blocks that, and the page itself as a last resort for reads. When a site rotates an id or moves a field, api-anything re-learns the request and only saves the new version if a replay succeeds.
A written step still needs a sandbox and still has not been tested against a real target, so this is a read tool today. The limits section is clear about that, and about the fact that a site challenging real Chrome still has to be cleared by hand. Even so, the shape here is useful: a saved operation is one more thing in a team's agent toolbox that costs the same next week as this week.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
