AI NewsOpen sourceAnnouncement
oc's authors say it cuts website tokens for agents 125-fold
oc is an MIT-licensed command-line tool that renders websites as compact text for AI agents, and its authors publish measurements putting the same 12 pages at 8,519 tokens through oc against 1,064,474 tokens as raw HTML.
Image: GitHub
Why it mattersWeb browsing is one of the largest token costs in an agent loop, so a tool that reports a measured 125-fold reduction on real pages changes what a research agent costs to run.
A project called oc, published under the only-cli organisation on GitHub, converts websites into compact command-line output built for AI agents to read. The repository was created on 18 August, carries an MIT licence and has 441 stars and 27 forks as of 2 September.
The pitch is a cost one. An agent that fetches a web page usually receives either raw HTML or a converted document, and both are large. oc renders the same page as terse structured text.
The numbers the project publishes
The repository reports a benchmark across 12 pages. Through oc those pages came to 8,519 tokens, against 1,064,474 tokens as raw HTML, which the authors describe as 125 times fewer. They report 15 times fewer tokens than Jina Reader and 59 times fewer than the accessibility snapshot produced by Playwright MCP.
Two individual pages are given. A YouTube page measured 345,487 raw tokens and 688 tokens through oc. The Node.js fs reference measured 275,425 tokens raw and 479 through oc.
The project also publishes cost comparisons against Claude's own tools. It puts a set of Wikipedia lookups at $0.23 through oc against $0.45 through WebSearch. For language documentation it reports 21% cheaper than WebFetch and 34% cheaper than WebSearch, and for dependency lookups 10% cheaper than WebFetch and 49% cheaper than WebSearch.
Every one of these figures is the project's own, measured by its authors on their own selection of pages, and nobody independent has repeated them. The comparisons against Jina Reader, Playwright MCP and Anthropic's tools are a project benchmarking itself against alternatives, which is the weakest kind of comparison there is.
How it lowers the numbers
oc includes shortcuts for more than 20 named sites, including Hacker News, Reddit, GitHub, Stack Overflow, Wikipedia, MDN, and the AWS, Python and Node.js documentation. A shortcut means the tool knows the structure of that site and can extract the part that carries meaning rather than guessing from the markup.
That is the real trade-off. The large reductions come from per-site knowledge, so they apply best to the sites someone has already written a shortcut for. A page outside that list falls back to general handling, and the project's headline ratio should not be assumed to hold there.
Per-site knowledge is where the savings come from
Browsing is expensive in a way that is easy to miss, because the cost arrives as context rather than as a line on a bill. An agent that reads five documentation pages during one task can spend more on those pages than on all of its reasoning.
The useful idea here is narrower than the tool. Sites an agent visits repeatedly, meaning your own documentation, your own admin pages and the two or three references your team relies on, are worth a purpose-built text view. The gains come from knowing the page, and a team knows its own pages better than any general extractor will.
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.
