Node HandbookNODE HANDBOOK中文
Clash Verge Rev · TUN

Clash Verge Rev: “reject loopback connection” with TUN on

The adapter exists, the core is running, and every node reads as failed. The log line that matters is a loopback connection being rejected — which describes the core's own outbound being caught by its own tun.

Node Handbook editorial team ·

Confirm this is the same failure

Local evidenceCheck firstDo not conclude
TUN never came up; a service or IPC start error appears firstService and core start stageThat this page's loopback diagnosis applies
TUN up, the same explicit proxy request works with TUN off, everything fails with it on, and the log shows reject loopback connectionEgress interface, routes, and whether DNS requests re-enter TUNThat all nodes are dead because the latency indicators are red
Only one node errors, or the explicit request fails with TUN off tooNode, rules, subscription or proxy portThat TUN is capturing itself
Only dns resolve failed, with no loopback line at the same timeDNS configuration and upstream request pathThat DNS proves a loopback problem

The report behind the symptom

Clash Verge Rev issue #8632 documents Windows 11 on client 2.5.8 with Mihomo 1.19.32. The log shows the TUN adapter listening, then connections initiated by verge-mihomo.exe recorded as reject loopback connection, alongside DNS resolution failures. With TUN off, explicit local-proxy requests from the same client recovered. The reporter closed the issue as completed on 16 November 2026.

Two limits matter. The explanation that “the core's outbound is being caught by its own tun” is the reporter's interpretation, not a packet capture or routing-table proof. And a closed issue does not mean a general fix was published. The reporter also noted that running the core from the command line with a minimal configuration, a replicated configuration, and under the SYSTEM account did not reproduce it — which is a reason not to claim a single root cause inside service mode.

Compare the same target with TUN on and off

  1. Record first. Client version, core version, Windows build, active network interface, the masked first error, TUN state, test time and selected node. Never publish the subscription URL, token or full configuration.
  2. Turn TUN off, keep the node and rule mode. Confirm plain networking, then make one explicit HTTPS request through the client's real HTTP or mixed port:
    curl.exe --proxy http://127.0.0.1:PORT -I https://example.com
    Replace the port with the value the client displays. A -I response only checks the target's response; it says nothing about download throughput or long-term stability.
  3. Turn TUN on and repeat the identical request. Record the core connection log at the same time. If it fails only with TUN on, compare process, destination address, matched rule, DNS error and the loopback rejection in time order.
  4. Check the physical egress interface and the system routes. Mihomo documents auto-route, auto-detect-interface and route-exclude-address as separate settings. Verify the runtime values before changing anything, change one item, and retest. Keep a copy of the configuration you can restore.
  5. If plain networking is still broken after TUN is off, stop the TUN experiment. Record the system-proxy switch, virtual interfaces, target routes and DNS state. Do not attribute it to “proxy residue” without evidence.

Why copying an exclusion list is a bad idea

An exclusion list encodes somebody else's interfaces, subnets and DNS addresses. Imported into your machine it can route your LAN, your DNS or your tunnel endpoint around the proxy — or worse, create the loop it was meant to prevent. If you need an exclusion, derive it from your own route table and interface list, and change one entry at a time.

What the reporter tried

Interface binding, disabling auto-detect and adding exclusion addresses were all attempted in that report without restoring every node. That result belongs to one environment. It is not a recipe: the same actions on a different interface layout can produce a different outcome, and the report does not establish which of them did what.

Speculation in the thread about third-party TUN conflicts is a suggestion from a collaborator. There is no evidence supporting uninstalling software, removing drivers or renaming DLLs as a response.

Scope and evidence limits

  • This sequence applies to a Windows case where you can obtain a before-and-after comparison around one TUN enable, and the log actually contains reject loopback connection.
  • Issue #8632 is a user account. This site has no packet capture, no route table and no reproduction on that network.
  • Different adapters, service modes, versions and DNS configurations can produce similar symptoms.
  • If configuration validation fails or the core never starts, use the service and startup path instead of adjusting TUN routes.

What to record

  • Whether TUN was up when the error occurred, and the first error in the log.
  • The result of the identical request with TUN on and off, including the port used.
  • The egress interface name and the default route at the time of failure.
  • Runtime values of auto-route, auto-detect-interface and any exclusion entries.
  • The order of the DNS error and the loopback rejection.
  • Whether the same failure occurs in another capture mode.