On October 8, Grass opened the Contents API. You give it a public URL. It returns HTML, Markdown, or plain text a model can read. You request access at lcr.net. There is no public price. Plans support 2 to 20 requests a second, and anything faster is a custom deployment.
This is a different product from the batches of training data that have carried the business. Chief executive Andrej Radonjic set the roadmap toward live retrieval around September 28, so an agent can read a page at the moment it needs that page. A corpus cut earlier misses a price, a filing, or a docs page that changed this morning. If you're building that agent, you care whether your URL comes back, and whether you can get access.
Grass is the Solana network Wynd Labs runs behind the GRASS token, built from home internet connections people share. Access is the gate. A demo that fetches a handful of pages can sit at 2 requests a second. An agent that walks a catalog will feel that rate on the first job, and 20 requests a second is the top of the published plans. Anything faster is a custom deployment, so get that ceiling in writing before you promise a product around the API. Then decide what you want back. HTML is the page as served. Markdown and plain text are what you hand a model when you want the words. You choose the URL. The call returns the page you named.
Grass's own test, run on October 1, compared the API with Firecrawl, Parallel, and Exa on 100 public URLs, with caching turned off where the other service allowed that.
| Provider | Full set | Hard domains |
|---|---|---|
| Grass | 97% | 98% |
| Firecrawl | 84% | Not broken out |
| Parallel | 76% | Not broken out |
| Exa | 66% | Not broken out |
A request can leave through a home ISP instead of a data-center address many sites challenge or block, fall back through its retrieval setup, and open a full browser when the page needs JavaScript, and nothing in the launch gives the API an on-chain job or a way for money to reach the people sharing bandwidth.
If you build agents that have to read live pages, request access and bring URLs you already fail to fetch from a data-center address. A storefront that throws a challenge, a job board that comes back empty, and a page that stays blank until JavaScript runs are worth the afternoon. Keep the text you get beside whatever your current fetch returns, and read it. A block page, a cookie wall, or an empty shell can look like a completed call if you only check that bytes came back. Do that before you let an agent act on the page. Your misses are the result that decides whether you build on this.
