Changelog

Behavior changes to the hosted service at api.webextract.dev (REST and MCP). Newest first.

2026-10-28

Statements in the docs, the home page FAQ and llms.txt that errors are not charged now say where they apply: REST calls, and MCP calls on the free tier or with a prepaid key. No API behavior changed.

The MCP, curl, Node and Python examples, the compare example, the benchmark data and the license text are now served from /examples/ (license: /examples/LICENSE.txt). Links on the home page and the benchmark page point there instead of to GitHub.

Failed fetches (for example HTTP 422) keep the same status code, error code and message, and add two fields: category (one of dns_failed, private_address, invalid_url, robots_disallowed, target_blocked, target_http_error, not_html, too_large, timeout, tls_error, redirect_error, target_unreachable, extract_failed, busy) and, when the target site answered with an HTTP error, upstream_status. charged: false is still included. Example: {"error":"upstream_error","message":"target site returned HTTP 404; not charged","charged":false,"category":"target_http_error","upstream_status":404}.
MCP tool errors keep the same first text block and add a second text block with the same fields as JSON: {"error","category","upstream_status"}. On the free tier or with a prepaid key, that block also includes "charged":false.
For failed calls, our access log now records only the target's domain and the stage where the fetch failed. It does not record the full URL, path or query string.
REST and MCP were both checked on the live service.

Pages with no readable text (empty markdown; for /v1/extract also no product and no tables) still return 200, but no longer use your free daily quota or prepaid credits. The response now includes warnings: ["empty_content"], a hint saying you can use the render endpoint instead, and charged: false. Applies to REST and MCP; both were checked on the live service.
The x402 paid path is unchanged: paid x402 calls that return an empty page are still charged at the quoted price.

/v1/extract tables now expand rowspan/colspan, so every row has the same number of cells and matches the Markdown output.

Server-rendered pages that already contain 150+ words of text are no longer flagged js_rendered_suspected; they now report no_product_markup. Genuinely empty app shells are still flagged.