Try
OUTBOUND checks run todaySince 2026-05-23

Exit & split-tunnel check

Measured by your browser under your own split rules: first whether you're on a VPN, a proxy or no proxy at all and how many lines you have, then the exit each site actually uses, whether IPv6 bypasses the proxy, and how that compares with each site's expected policy.

Reading the results

What this page checks and how to read it

Split tunneling sends different traffic through different exits. This page reads the direct and fallback exits with three reference probes, then reads an echo on each site's own domain to see which exit it really uses — every request is sent by your browser through your own rules.

What is being checked

Four things: whether direct and proxy really are two exits; whether each site's measured exit matches what's expected (ChatGPT through the proxy, iQIYI direct); whether IPv6 bypasses the proxy; and whether each site is reachable.

A site's exit can only be measured when its own domain has a readable echo (usually Cloudflare's /cdn-cgi/trace); those sites are marked "Measured". Others are only tested for reachability, marked "Inferred" or "Unknown", and never placed under a specific exit. IP quality isn't assessed here — open an exit IP for its full report.

How the connection type is judged

HTTP exits only show whether a split exists; whether it's a VPN or a proxy, and whether UDP and DNS are covered, comes from two more checks.

  • WebRTC (UDP)UDP is sent to Google's and Cloudflare's STUN servers and the reflected public address is read: leaving through the proxy means UDP is covered (TUN / VPN); leaving through the direct exit means only HTTP is proxied (system or browser proxy) and WebRTC reveals your real address.
  • DNS probeThe browser resolves one-off subdomains and our authoritative NS records the recursive resolver: on the direct network it's local resolution, on the node's network it's fake-ip or remote resolution, and public DNS is listed separately.
  • Echo flagswarp, gateway and loc=T1 in the Cloudflare echo directly mark WARP, Zero Trust and Tor.
  • Tunnel or relayThe server observation point (fp.cleanip.io:8443) terminates TCP itself, captures the first SYN's TTL and option order and compares them with your system: Windows sends no timestamps and TTL 128; macOS / iOS send mss,nop,ws,nop,nop,ts,sackOK; Linux sends mss,sackOK,ts,nop,ws. Seeing the node system's stack means TCP is relayed (L4); seeing your own machine's stack means a tunnel (L3). Linux / Android users meeting a Linux node can't be told apart that way, so the gap between its RTT to the node and the browser's round trip decides instead. System and browser proxies both leave UDP alone, so they aren't told apart here.

Reference probes

Each kind of rule gets its own probe; comparing them shows which kind isn't working.

  • IP-rule directEcho from the China IP address 124.222.230.137 — no domain rules involved, it hits GEOIP,CN / IP-CIDR.
  • Domain-rule directEcho from iQIYI's data.video.iqiyi.com, which hits geosite:cn style China domain rules. With only a GEOIP rule, it goes through the proxy.
  • Fallback (MATCH)Echo from v4.cleanip.io, normally on no list, so it falls through to the final rule; if you wrote a rule for cleanip.io yourself, it measures that rule instead. In blacklist mode it going direct is normal.
  • IPv6v6.cleanip.io has only an AAAA record and reads the exit your IPv6 traffic uses. On the direct network while the fallback is proxied means IPv6 bypasses the proxy.
  • Google (approximate)Google sites have no readable echo, so the page reads the subnet (/24) Google Public DNS sees and compares it with each exit's subnet — an approximation only.
  • TCP observation pointEcho on fp.cleanip.io:8443, on the same line as the fallback; it bypasses the CDN and terminates TCP and TLS itself, so it can return the server RTT, peer MSS / window scale / timestamps and JA4, and its response carries Timing-Allow-Origin so the browser can read the connect / TLS / first-byte breakdown. If port 8443 is blocked, only this piece of evidence is missing.
  • HTTP/3 observation pointfp.cleanip.io:8444 listens on TCP and UDP at once: a TCP reply carries Alt-Svc, then the browser tries QUIC. If one request goes over HTTP/3, UDP passes along this line (TUN / VPN covers UDP, or the node relays UDP); if it stays TCP only, UDP is blocked or the proxy doesn't relay UDP. It cross-checks WebRTC and still judges UDP when WebRTC is disabled.

How to read the verdict

The verdict combines the three reference probes with every measured site: it first finds the direct exit, then checks whether each measured site went direct or through the proxy as expected. Any mismatch rules out "Split routing works".

  • Split routing worksDirect exit in China, proxy exit abroad, and every measured site matches. A direct fallback (blacklist mode) can be a normal result and is noted separately.
  • Split in place, with mismatchesA site that should use the proxy went direct, or one that should go direct was proxied, or one kind of direct rule isn't working. Check the rules against "Mismatches and fixes".
  • IPv6 bypasses the proxyIPv4 follows your rules but IPv6 leaves through the direct network, so sites with AAAA records (ChatGPT, Claude…) go direct.
  • Global proxyThe direct reference probes and the fallback land on the same datacenter exit outside mainland China: most likely a global proxy (China traffic is proxied too), but your direct network may itself be a datacenter — the two can't be told apart. A single datacenter exit in mainland China is reported as a single exit, since a proxy there can't reach blocked sites.
  • No proxyThe direct reference probes and the fallback land on the same access-network (home broadband or mobile) exit with no sign of a proxy; outside mainland China it may also be a global proxy through a residential node, which can't be told apart.
  • Exit location unknownWhen the location or the direct exit's network type can't be looked up, only IPs and ASNs are compared; if it can't be confirmed which exit is direct, expectations aren't compared and nothing is reported as matching.
  • Reference probes unreachableIf neither direct reference probe answers, the direct exit can't be determined — common outside China or when your proxy REJECTs the China echo. If the fallback probe fails and no site reaches a proxy exit (the proxy node may be down), nothing can be concluded either.

What the site columns mean

Grouped by category by default; switch to grouping by exit.

  • Exit"Measured" shows the exit read from the site's echo and its role (direct / proxy); when three requests came back from different exits, the number of rotating exits is shown; rows that differ from the expected route are highlighted.
  • ReachabilityAn HTTPS request from your browser, not an ICMP ping. A site only counts as unreachable when no HTTP response comes back at all (reset, DNS failure, certificate error, timeout).
  • FamilyShown for measured sites only: an IPv6 address in the echo means that connection used IPv6.
  • Resolved IPWith "Show resolved site IPs" ticked, sites are resolved over DoH: China sites through AliDNS with your direct exit's /24 (EDNS Client Subnet); others through Cloudflare, falling back to Google, without a subnet. It's only the DNS answer, not the address your browser actually connects to.

Common results and fixes

The most common patterns and where to look; "Suggested rules" exports them in your client's format.

  • China domains go through the proxyChina domain rules are missing. Add geosite:cn or a China domain rule set as DIRECT — essential in fake-ip mode.
  • IP-rule direct goes through the proxyGEOIP,CN,DIRECT is missing or sits after the fallback rule.
  • A should-proxy site goes directThe domain isn't on your proxy list (common in blacklist mode). Add it to a proxy rule or switch to whitelist mode.
  • IPv6 bypasses the proxyLet your client handle IPv6 (enable IPv6 in TUN mode), or turn off IPv6 on the system.
  • A few sites are unreachableThe site is blocked, the node is down, or a rule REJECTs it. In "Suggested rules", copy the "Should go direct" / "Should use the proxy" groups and check them one by one.

Why DNS answers differ from other tools

Sites like Taobao, QQ and Google use geo-DNS: the authoritative server answers with a node near whoever asks, so there is no single correct IP.

China sites are resolved with your direct exit's subnet attached, and the resolver and every record are listed on hover; when comparing with another tool, note where it asks from and through which resolver.

Data sources and privacy

Every probe is sent straight from your browser, so each party sees the exit IP you use on that route. Results aren't uploaded; your site list stays in this browser.

This page contacts:

  • CleanIPThree echoes (124.222.230.137, v4 / v6.cleanip.io), the TCP observation point (fp.cleanip.io:8443, which records only this connection's TCP facts and first SYN and keeps no history) and the HTTP/3 observation point (fp.cleanip.io:8444), plus lookups of each exit IP's location and network type: direct IPs are checked against the local database only (no paid sources, no profile written); only exits judged to be proxies, classified as data centers by the local database and located in a different country from the direct exit get one full lookup; residential, unknown-type or same-country proxy exits use the local database only, and when the direct exit itself is in a data center every exit uses the local database only. When the direct references and the fallback land on the same data-center exit outside mainland China (including its same-network IPv6), a global proxy cannot be told from a data-center line and those addresses get a full lookup.
  • Tested sitesEvery site on the list (including the iQIYI, Toutiao, Tencent GDT, Cloudflare China and ip.cn echoes and Google Public DNS), plus any custom domains you add.
  • DoH resolversOnly when "Show resolved site IPs" is ticked: AliDNS (China sites, with your direct exit's /24), Cloudflare and Google (other sites, no subnet).
  • favicon.laIcons for preset sites without a local icon; custom domains are never sent to it.
mimetic
mimetic
mimetic