Highlights
A call that can't get the page should say so fast. Auto now learns, site by site, whether searching more exits still brings the page back, and where it almost never does, the call ends at the first empty search. Single, Proxy Finder and Auto also answer a clear 400 for a URL with a space inside it, revoking a session in the Dashboard signs that device out at once, and MCP 0.8.4 keeps offload_large working after its payload folder is cleared.
What's New
Auto learns where more exits won't help
Last week Auto started keeping to three browser requests a call on sites where the ones past the third almost never bring back the page. Now it learns the next step as well. If searching for an exit after those three browsers hasn't delivered on that site lately either, the call ends at the first search that comes back empty. That holds for the next 24 hours, and then Auto checks again.
The answer says what happened: error starts with the browser has not got through to this site, and status holds the site's last answer. So you know why the call ended before you decide whether to retry.
The stop is set per site, not per API key. Because it applies to everyone calling that site, Auto judges it by the site's own answer and not by your validate rule (a strict rule of yours can't switch it on). And a call with exitClass: "premium" still gets its premium exits.
Auto's job is finding the route. Once it hands you a working session on a site, replay that session on Single or Browser for the volume. Smart Fetch (Auto) shows the replay and lists every way an Auto call can end early.
A clear 400 for a space inside a URL
Crawlers pick up links exactly as a page writes them, and pages write odd things: javascript: void(0), a share link with a whole sentence in it. Send Single, Proxy Finder or Auto a URL with a space or a control character inside it, and you now get a 400 before the request leaves FourA:
invalid URL: it contains a space; encode it as %20 (the URL as sent: "...")
Nothing is fetched, no credits are spent, and the outcome is client_error (see Request Outcomes). Browser is the exception: it encodes the space itself and fetches the page.
The fix on your side takes a few lines. Resolve each link against the page it came from, skip the ones that aren't pages, and encode the rest:
from urllib.parse import urljoin, quote
import requests
SKIP = ("javascript:", "mailto:", "tel:")
def to_fetch(page_url, href):
href = href.strip()
if href.lower().startswith(SKIP):
return None
# %20 for a space; an already-encoded link stays as it is
return quote(urljoin(page_url, href), safe=":/?#[]@!$&'()*+,;=%")
url = to_fetch("https://example.com/shop/", "/items/blue shoes?size=42")
# -> https://example.com/items/blue%20shoes?size=42
r = requests.post(
"https://api.foura.ai/api/single",
headers={"X-API-Key": "pk_live_..."},
json={"url": url},
).json()
Common Issues has the symptom and the fix side by side.
Revoking a session signs the device out at once
Settings has a Sessions tab that lists every session across your devices. Pick sessions there, click Revoke Selected, and those devices are signed out of the Dashboard straight away, the same as with Sign Out Others. A device you've revoked has to sign in again before it can open your Dashboard.
It's the button for the laptop you left signed in at a client's office. One device goes, and every other session of yours stays as it is.
MCP 0.8.4 keeps offload_large returning its link
With offload_large: true, the MCP server writes a response of 50 KB or more to disk and hands your client a resource_link in place of the body. On your own instance, the docs ask you to clear files older than an hour from the payload directory. In 0.8.4, offload_large keeps returning its link after that cleanup (or after the system empties its temp folder), because the server creates the folder again before every write. A payload whose files are gone reads as Payload not found.
The release refreshes locked dependencies too, clearing published security advisories, and npm audit reports no known vulnerabilities. The tools, their schemas and the requests they send don't change. Update @fouradata/mcp, or point your client at the hosted server, which already runs 0.8.4. Setup and every option are in the MCP Server docs.
The cheapest failed call is the one that ends in seconds and tells you which line to change.