Why Claude feels flaky before you blame the model

In 2026, Claude is a daily driver for writing, coding support, and office automation. When the Claude web interface sits on an endless loading ring, the first instinct is to assume Anthropic is down or throttling you. Sometimes that is true. Far more often, especially on networks where cross-border paths are uneven, the browser is waiting on a chain of requests where one hop is only half-routed. A primary document loads, a script bundle arrives, then a long-lived connection to an API host stalls because your profile silently matched DIRECT or a congested catch-all proxy group.

This article stays in the lane of consumer and professional AI tool proxy routing. It complements our Grok and xAI split-routing guide and the Chrome Gemini and Google routing walkthrough, but the hostnames, auth flows, and CDN edges are not interchangeable. The goal is simple: make Anthropic traffic boring again by giving it explicit Clash split rules instead of hoping your default policy group happens to be healthy tonight.

If you are new to policy groups and ordered rules, start with the beginner routing tutorial on this site, then return here for the vendor-specific checklist. Everything below assumes a Mihomo-class core, a GUI or dashboard that exposes logs, and permission to edit your own YAML or merged profile.

What Anthropic-shaped traffic looks like in practice

Modern chat products are not single-URL monoliths. The Claude experience pulls HTML and JavaScript from a primary site, opens streaming channels for tokens, calls REST or WebSocket surfaces on related hosts, and occasionally reaches authentication, billing, telemetry, or support widgets on third-party domains. For planning proxy rules at the domain level, treat the product as a small bundle of hostnames rather than one memorable entry point.

At the time of writing, public clients routinely touch claude.ai for the consumer web shell and anthropic.com for corporate marketing, documentation, and account flows. API and console traffic may appear on subdomains such as api.anthropic.com or console.anthropic.com, and those names can shift as vendors rebalance infrastructure. Your most reliable source of truth is your browser developer tools: open the Network panel, filter by domain, and compare a working full-tunnel session with a flaky split-tunnel session. Anything that appears only in the broken profile belongs in your allow list ahead of broad community lists.

Because most flows use HTTPS, the first meaningful identifier your local proxy often observes is the TLS Server Name Indication string. DOMAIN and DOMAIN-SUFFIX rules align naturally with SNI for ordinary TCP TLS. When the stack negotiates HTTP/3 over QUIC, UDP-shaped packets may follow a different path through your node unless UDP is supported end to end. That distinction matters when the UI symptom is intermittent: everything works for five minutes, then streaming stalls until you refresh.

Give Anthropic traffic its own policy group

A common anti-pattern is dumping every foreign site into one oversized PROXY selector and rotating nodes when anything misbehaves. That obscures the real failure mode: a single unhealthy exit, a rate limit on one region, or a UDP-partial relay masquerading as a general outage. Claude access unstable symptoms are especially sensitive to latency variance because the UI batches many small requests and keeps connections warm.

Create a dedicated policy group—for example ANTHROPIC or AI_CLAUDE—with two or three nodes you trust for low packet loss, not the same pool you use for bulk downloads. If your client supports URL tests or latency checks, wire them to real endpoints rather than generic ping targets. The objective is isolation: when the web app misbehaves, you switch only that group, read its logs, and compare against a neutral benchmark site.

If you already maintain region-specific groups, decide whether Anthropic should follow your United States selection, a low-latency anycast group, or a specialty relay offered by your operator. The answer depends on peering and congestion, not on branding. Revisit the decision when you change providers, because AI vendors migrate CDNs faster than most home users update subscriptions.

Domain rules, rule-sets, and the order that actually matters

Mihomo-family cores evaluate rules from top to bottom until the first match wins. That sounds trivial until a broad GEOIP line or a community RULE-SET labeled for domestic direct routing consumes a hostname you mentally classified as foreign. The page skeleton loads, then a secondary call to a CDN edge with a different IP classification stalls, producing the familiar spinner that looks like application bugs.

Place explicit DOMAIN-SUFFIX entries for claude.ai and anthropic.com, plus any additional names you captured from devtools, above those broad strokes. If you rely on remote rule-set providers, keep a thin inline “priority strip” for vendors you depend on daily, then stack community lists underneath. Our advanced rule providers article explains how RULE-SET ordering interacts with GEOIP and custom lists; the same discipline applies to Anthropic traffic.

Use DOMAIN-KEYWORD sparingly. Keywords that match innocent substrings can steer unrelated telemetry into your AI group or bypass security lists you believed were global. Prefer suffix rules for owned domains, then verify with live logs that the intended policy group is selected. If you maintain separate stacks for API clients versus browsers, duplicate the suffix list in both profiles or import a shared snippet to avoid drift.

When a subscription merge injects thousands of lines you did not read, assume nothing about order until you inspect the final runtime configuration. Merge tools can reorder sections in surprising ways. After any update, confirm that your Anthropic strip still sits ahead of catch-all regional rules.

DNS: the hidden reason “the site works on phone but not on Wi-Fi”

Misunderstood DNS generates more phantom outages than any single proxy protocol. If your ISP resolver returns split-horizon answers, aggressive TTL trimming, or different addresses from a public resolver, the browser and the tunnel may not even agree on which endpoint they are contacting. Align nameserver policy with your routing story: either resolve foreign names through the tunnel pipeline or accept that fake-ip mode changes how domain rules interact with client-side caches.

With fake-ip enabled, newcomers are sometimes surprised when a DOMAIN rule behaves differently than it did on a classic resolver setup. The client and core split resolution responsibilities, and stale OS caches can exaggerate the effect. When Claude works on cellular but not on home Wi-Fi, compare DNS profiles before you swap exit nodes. A disciplined test is to query the same hostname through the proxy DNS path and through the ISP path side by side; if answers diverge materially, fix DNS first.

Watch DNS over HTTPS inside the browser. A profile that intercepts only system DNS may miss encrypted resolver traffic unless TUN captures it. If you intentionally avoid TUN, configure the browser to use system DNS or expect occasional holes where the application quietly bypasses your assumptions. For a full map of fake-ip pitfalls, the Clash official home page links into documentation that pairs well with hands-on testing.

QUIC, UDP, and HTTP/3: when the UI “almost works”

When properties enable HTTP/3, browsers may prefer QUIC over TCP. If your outbound path drops UDP, deprioritizes it, or handles it on a different chain than TCP, you will see intermittent stalls that resemble product bugs. Some nodes advertise UDP support unevenly across regions or ports, which shows up as “first message succeeds, streaming freezes.”

Structured mitigations include temporarily disabling QUIC in the browser to compare behavior, confirming that your Anthropic policy group forwards UDP when you require full HTTP/3 through the proxy, and reading client documentation on how underlying protocols map to proxy-groups. This is not a recommendation to disable HTTP/3 forever; it is a diagnostic lens that tells you whether your stack is QUIC-complete.

WebRTC and other UDP-first features can amplify the same class of problems. If your threat model allows, test in a conservative browser profile while you validate routing, then re-enable experimental features after you have a stable baseline.

Reading logs without drowning in noise

Logs are the honest ledger of rule priority. When you reproduce a single failing action—opening a new chat, uploading a file, or resuming a long session—watch for the hostname, the matched rule type, and the selected policy. If you expected ANTHROPIC but see DIRECT, you do not have a Claude problem; you have an ordering problem or a stale merge artifact.

Enable sufficient verbosity to print rule names or providers when your client supports it. Correlate timestamps with browser network rows so you know which request corresponds to which log line. If logs show repeated TLS handshake timeouts on one node but not another, promote the healthier member within the Anthropic group rather than toggling global modes.

When logs implicate a CDN edge with a cryptic hostname, resist the urge to keyword-match blindly. Capture the exact suffix, confirm ownership with a quick lookup, and add a precise DOMAIN-SUFFIX entry. Over-broad keyword rules age poorly and create security debt.

Common misroutes and how they feel in the Claude web app

What you see What to inspect first
Spinner on first paint while other sites work An early DIRECT rule or GEOIP match on a CDN name you forgot to suffix-list
Login succeeds, conversation never starts Secondary API host still on DIRECT; WebSocket path blocked or routed through a UDP-partial hop
Captcha or risk challenges loop Mixed egress IP across tabs; align cookies with a consistent Anthropic group selection
Works on one network only Resolver differences, captive portals, or IPv6 leaks outside the tunnel
Keep a three-host smoke test: one domestic site that must stay DIRECT, one generic foreign site that must use your default proxy, and one Anthropic hostname that must hit ANTHROPIC. After every DNS or rule change, verify all three before you reopen the Claude tab.

Self-check checklist before you open a support ticket

  1. Confirm the client loaded the merged profile without silent errors.
  2. Verify Anthropic-related hostnames appear in logs with the intended policy group, not accidental DIRECT.
  3. Match DNS modes between working and failing scenarios; test for DoH leakage.
  4. Run TCP and UDP checks against the same node; compare QUIC enabled versus disabled.
  5. Validate system time and certificate trust if TLS errors look nonsensical.
  6. Retry in a clean browser profile to discard stale service workers or extensions.

Ethics, terms of service, and realistic limits

Routing traffic through another region can conflict with terms of service or local regulation. This guide explains mechanics for readers who are legally permitted to use Anthropic products from the location implied by their routing choices. It does not encourage bypassing fraud prevention, abuse monitoring, or payment controls. If the service presents a challenge screen, treat it as a risk signal rather than an invitation to hammer retries through random exits.

Technology cannot fix every upstream incident. When Anthropic endpoints are genuinely degraded, patience and official status channels beat configuration churn. What a disciplined Clash setup gives you is confidence: you eliminated self-inflicted routing noise before concluding the product is unusable.

Closing: predictable paths beat hopeful defaults

Compared with toggling one global switch and hoping for the best, explicit Clash Anthropic split routing trades a handful of deliberate lines for quieter debugging. You get a named policy group, a short domain list ahead of community rule-sets, DNS that matches your tunnel narrative, and a practical lens on UDP-heavy HTTP stacks. That predictability is why teams standardize on Mihomo-class cores in the first place: fewer mystery spinners, clearer logs, and faster recovery when vendors shift edges overnight.

If you already split Grok or Google traffic, add Anthropic the same way rather than mixing unrelated AI vendors into one overloaded selector. The cognitive cost of three well-named groups is lower than the cognitive cost of one “AI” bucket that fails for different reasons every week.

When you are ready to align the client experience with the same clarity, the Clash download page lists installers and updates in one place—distribution first, transparency about source second.

→ Download Clash for free and experience the difference

Download Clash from this site