1. Scope: a research workstation is not just a browser
Academic work creates a particularly mixed networking profile. A single literature review may begin in Google Scholar, continue through a publisher platform, open a university proxy login, download a PDF, send metadata to Zotero, and finish with a collaborative edit in Overleaf. At the same time, the same computer may need direct access to a campus intranet, a laboratory instrument, a local Git server, a departmental VPN, or a library catalogue that only works from an institutional address. Treating this entire workload as one “international” or “domestic” category is the fastest way to create confusing failures.
This guide uses Clash or a Mihomo-compatible client to build split routing around observed research traffic. The objective is not to proxy every request. It is to give services that genuinely need a particular network path a predictable policy group, while keeping campus systems, private repositories, local printers, remote desktop tools, and ordinary low-latency traffic on DIRECT or on the route required by your institution.
The examples are intentionally conservative. Academic publishers change content-delivery hosts, Zotero libraries may use WebDAV or an institution-managed storage service, and Overleaf can load assets from several related domains. Use the examples as a starting point, then confirm the actual destinations in Clash logs and browser developer tools. A rule copied from a forum is a hypothesis, not proof that it belongs in your profile.
2. Map the complete workflow before writing rules
Start with a short inventory of what happens during a real research session. Do not test only the homepage of Scholar and assume that the rest of the workflow shares its network behavior. Open a fresh browser window, search for a paper, follow one DOI link, authenticate through the library if necessary, download a document, add it to Zotero, synchronize the library, and open an Overleaf project. Record which steps succeed and which fail. Repeat the process once with Clash disabled and once with the intended policy group selected.
The inventory should include both visible applications and background processes. Chrome, Firefox, or Safari may follow the operating system proxy, while Zotero uses its own networking stack. A PDF viewer may fetch supplementary material independently. Git, a LaTeX package manager, and an editor extension may use command-line HTTP clients that ignore desktop proxy settings unless environment variables are present. Overleaf collaboration can involve browser WebSockets, upload requests, and connections to project or compilation services that do not appear when you merely load the landing page.
Separate traffic into useful categories
Four categories are usually enough for a first design. The first is research access: Scholar, publisher pages, DOI services, open-access repositories, and documentation that you intentionally want to reach through a stable research exit. The second is library and identity access: university authentication, EZproxy, Shibboleth, SAML, Microsoft Entra, or another institutional identity provider. This category often must remain on a campus VPN or direct route, so do not automatically place it with public research domains. The third is collaboration and storage: Overleaf, Git hosting, cloud drives, Zotero sync, WebDAV, and any private project endpoint. The fourth is local and protected traffic: RFC1918 networks, campus hostnames, laboratory devices, internal databases, and services covered by an institutional policy.
Naming these categories prevents a common mistake: using one broad rule such as DOMAIN-SUFFIX,edu,RESEARCH. Educational domains contain public libraries, private campuses, login services, and unrelated sites. A suffix that looks convenient can send an internal administrative portal through an external relay or can make a library authentication redirect fail. Specificity is more valuable than a large list.
3. Design policy groups for research tasks
Create policy groups that describe intent rather than geography. A group named RESEARCH can contain reliable nodes suitable for reading and downloading academic material. A separate COLLABORATION group can serve Overleaf, Git hosting, and remote storage if those services perform better through a different exit. Keep DIRECT available for campus services and add a CAMPUS_VPN group only when your Clash client and institutional VPN can coexist safely.
The group should also express acceptable failure behavior. If a research node becomes unavailable, silently falling back to DIRECT may create a misleading result: the page loads partially, a search returns fewer records, or an authentication request is sent from an unexpected region. For sensitive workflows, a deliberate failure is easier to diagnose than an invisible fallback. For ordinary public documentation, direct fallback may be reasonable. Decide this per group instead of inheriting a default without reviewing it.
Choose stable exits instead of merely fast exits
Literature work is not measured only by the first page’s ping time. A research session may download several large PDFs, maintain a long browser login, or upload project assets to Overleaf. A node that wins a short latency test can still have poor sustained throughput, unstable TLS connections, or an exit reputation that triggers repeated verification pages. Use a url-test group if its health checks reflect the services you care about, and compare real download and login behavior before changing nodes every few minutes.
For collaborative writing, stability matters more than a small difference in round-trip time. A dropped WebSocket can interrupt presence indicators or make a compile result appear stale. If Overleaf is reliable through one node but Scholar performs better through another, use two groups rather than forcing unrelated traffic through a single compromise.
4. Observe hostnames from Clash logs and browser tools
Turn on enough logging to see policy decisions without collecting unnecessary personal data. In Clash Verge Rev, reproduce one action at a time and watch which domain is matched, which rule is selected, and which group handles the connection. In a browser, the Network panel can show document, script, XHR, fetch, and WebSocket destinations. Export only the hostnames and rule outcomes you need for configuration. Do not publish cookies, authorization headers, full URLs containing private query parameters, or downloaded paper metadata in a public issue.
Begin with exact observations such as scholar.google.com, doi.org, a known repository hostname, and the actual Overleaf domain used by your account. Static assets may come from gstatic.com or another CDN, but adding an entire provider suffix before you understand the traffic can create a wide blast radius. Zotero may contact storage, sync, citation-style, update, or translator endpoints at different times. Test a library sync, not just application startup, and note every endpoint that receives a connection during that operation.
Keep an endpoint notebook
A small table in your research team’s internal documentation is more useful than a permanent comment saying “needed for Zotero.” Record the date, application version, operating system, action tested, hostname, selected rule, and result. Mark each entry as confirmed, suspected, or obsolete. This makes future maintenance safer when a service migrates to a new CDN or when a university changes its authentication provider.
| Workflow | First endpoints to observe | Validation action |
|---|---|---|
| Academic search | scholar.google.com, DOI and repository hosts |
Run a search and open two different result links. |
| Zotero | Sync, storage, WebDAV, translator, and login hosts | Sync a test collection and verify attachment availability. |
| Overleaf | Project, asset, WebSocket, and compilation hosts | Edit a file, compile, and download the generated PDF. |
| Campus services | Library proxy, SSO, intranet, and private DNS names | Authenticate and open an internal resource without the public relay. |
5. Write specific rules before broad rules
Mihomo evaluates rules from top to bottom, so ordering is part of the configuration’s meaning. Put private networks and explicitly protected institutional domains first. Follow them with measured research and collaboration domains. Place geographic shortcuts after the specific service rules, and keep the final MATCH rule visible and intentional. If a broad rule appears first, a later exact domain rule will never be reached.
The following scaffold is illustrative. Replace the policy-group names and hostnames with the values that your own logs confirm. The googleapis.com and gstatic.com suffixes are deliberately shown as optional examples; a narrow configuration is usually easier to audit than a broad one.
# Example groups: RESEARCH, COLLABORATION, DIRECT
rules:
- DOMAIN-SUFFIX,lab.example.edu,DIRECT
- DOMAIN-SUFFIX,internal.example.edu,DIRECT
- IP-CIDR,10.0.0.0/8,DIRECT,no-resolve
- IP-CIDR,172.16.0.0/12,DIRECT,no-resolve
- IP-CIDR,192.168.0.0/16,DIRECT,no-resolve
- DOMAIN,scholar.google.com,RESEARCH
- DOMAIN,doi.org,RESEARCH
- DOMAIN-SUFFIX,overleaf.com,COLLABORATION
- DOMAIN-SUFFIX,zotero.org,COLLABORATION
- DOMAIN-SUFFIX,webdav.example.edu,DIRECT
- DOMAIN-SUFFIX,gstatic.com,RESEARCH
- GEOIP,CN,DIRECT
- MATCH,DIRECT
Do not copy the broad suffixes blindly. If a particular static asset is required, add its observed hostname and test the workflow again. If your institution owns an Overleaf deployment or a Zotero WebDAV endpoint, place that exact domain above any public-service rule. The same principle applies to a university’s SSO host: authentication may need the campus route even when the public article page uses the research route.
Domain rules versus process rules
Domain rules are usually the best first choice because they remain understandable across operating systems. Process rules can help when a command-line client connects to an IP address, uses an unexpected SNI, or bypasses normal proxy environment variables. However, process names are easy to misidentify. Zotero can spawn helper processes, and a browser may create several children with different names. A process rule that captures the parent but misses the child gives a false sense of coverage. Use it only after connection logs or operating-system tools show that domain rules and system proxy settings are insufficient.
6. Handle DNS, fake IP, and TLS carefully
DNS behavior is often the hidden difference between “the rule looks correct” and “the application still cannot connect.” In fake-IP mode, Clash returns synthetic addresses and later maps them back to domain names. This can work well for ordinary browser traffic, but institutional authentication systems, local discovery, and applications that perform their own DNS resolution may behave differently. If a university hostname resolves inside campus DNS but not through a public resolver, ensure that your DNS strategy can reach it before routing the connection.
Start by confirming whether the failing hostname is resolved by Clash, by the operating system, or by the application itself. A browser may appear healthy while Zotero uses a separate resolver. Check for inconsistent IPv4 and IPv6 paths, especially when a campus VPN advertises private IPv6 routes. A connection that alternates between a local IPv6 address and a proxied IPv4 address can produce intermittent login or download failures.
TLS inspection is another boundary. Do not install a custom certificate or enable interception merely because a PDF download failed. First determine whether the failure is DNS, routing, authentication, or a server-side access restriction. Many publishers and identity providers expect normal certificate validation and may reject unusual interception behavior. Keep the client’s certificate store current, and never disable certificate verification as a shortcut.
7. System proxy versus TUN mode for research applications
System proxy mode is the most transparent starting point. Configure the Clash mixed port, enable the operating system proxy, and verify that the browser follows the intended group. This approach keeps the blast radius small and makes it easier to leave campus VPN software, local discovery, and protected administrative applications alone. It is also easier to explain to a supervisor or security reviewer because only proxy-aware applications participate.
The limitation is that not every application honors the system proxy. Zotero, a command-line downloader, a LaTeX package manager, or an editor extension may use a separate HTTP library. Shell tools often need explicit variables such as HTTP_PROXY, HTTPS_PROXY, and NO_PROXY. A useful NO_PROXY value can include localhost, private address ranges, campus domains, and internal service names, but confirm your institution’s exact requirements.
TUN mode captures traffic at the routing layer and can solve stubborn proxy bypasses. It also affects more traffic: SSH sessions, printers, virtual machines, campus VPN adapters, container networks, and local discovery may all change behavior. Enable it only after a reproducible test demonstrates that system proxy mode cannot handle the required application. When TUN is active, document exclusions and test sleep, wake, VPN reconnect, and network changes. A configuration that works at home may conflict with a university VPN on the next morning’s campus network.
Make command-line tools explicit
A research workflow often includes Git, Python scripts, citation exporters, and LaTeX package downloads. Set proxy variables in the environment where the command actually runs, not only in a graphical client. Then test the command with verbose output and inspect its connection in Clash. Keep credentials out of shell history and avoid putting proxy passwords directly into shared project files.
# Example for a shell session; adjust the local port to your client
export HTTP_PROXY=http://127.0.0.1:7890
export HTTPS_PROXY=http://127.0.0.1:7890
export NO_PROXY=localhost,127.0.0.1,.internal.example.edu
# Test a public endpoint without exposing credentials
curl -I https://scholar.google.com
curl -I https://www.overleaf.com
If a tool uses a SOCKS interface instead, configure its documented SOCKS option rather than assuming that an HTTP proxy URL will work. If the tool ignores both, TUN may be appropriate, but verify the process and destination in the Clash log before changing the entire workstation’s routing model.
8. Zotero-specific checks
Zotero is a desktop application, so a successful browser test does not prove that synchronization will work. First separate the functions you use: metadata lookup, translator updates, library synchronization, attachment storage, WebDAV, and institutional login. A library may sync metadata successfully while attachments fail because they use another storage endpoint. Conversely, attachments may be available while a translator cannot fetch a page because the browser and Zotero use different proxy settings.
Test with a small collection and one disposable attachment. Watch the Clash log while forcing a sync, then repeat after restarting Zotero. This reveals whether the application reads the proxy configuration at startup and whether a helper process creates new connections later. If your group uses WebDAV, treat the WebDAV hostname as a separate service and keep it on the institutional route when required. Do not assume that a domain containing “zotero” represents every endpoint used by your configured storage provider.
When a sync error appears, record the exact time, operation, and endpoint before changing multiple rules. Check system time, certificate errors, account status, storage quota, and server response separately. A proxy rule cannot fix an expired account token or a full WebDAV quota. Small, reversible rule changes make it possible to identify the real cause.
9. Overleaf and collaborative writing
Overleaf combines ordinary page requests with active collaboration, file uploads, compilation, and generated artifacts. A page that opens successfully may still fail when you type, compile, or download a PDF. Test each operation: open a project, edit a harmless comment, save, start a compile, inspect the log, and download the result. If presence indicators or live updates stop working, investigate WebSocket handling and the selected policy group rather than only measuring page-load latency.
Long-lived connections benefit from a stable exit and a relay that handles the transport used by your client. Avoid switching nodes in the middle of a compile or while uploading a large figure. If your profile uses automatic selection, choose a sensible interval and ensure that a health check does not cause frequent changes during active collaboration. When multiple authors share a project, consistency is valuable: a rule that works only from one country or one ISP may create confusing differences in compile and synchronization behavior.
Overleaf projects may also pull packages or figures from external locations, depending on how the document is structured. Keep the project’s source reproducible and avoid relying on a browser-only proxy exception. If compilation occurs remotely, your local Clash configuration affects the project’s web session and upload path, not every outbound request made by the remote compiler. Diagnose those layers separately.
10. Protect campus services and institutional access
The most important exception list is often the one you do not see in a public proxy guide. Add internal suffixes, private address ranges, library proxy domains, laboratory networks, and identity-provider hosts according to your institution’s documentation. Put these rules before broad research rules. If the campus uses split DNS, preserve the resolver path required for internal names; routing a private hostname through an external DNS server can fail before Clash has a chance to select a policy.
Be careful with SSO redirects. A public publisher may redirect to a university identity provider, which then redirects back to the publisher. If one leg uses a research relay and another requires the campus VPN, the login loop can look like an account problem. Observe the entire redirect chain and assign each domain according to institutional guidance. Never solve a loop by weakening certificate checks or by adding every identity provider suffix to a public proxy group.
Also test ordinary work that should remain ordinary: SSH to a lab server, access to a local printer, a departmental Git repository, and a campus file share. Researchers often notice only when Scholar works, then discover later that a broad rule broke a scheduled backup or instrument connection. A good configuration improves the target workflow without silently damaging adjacent work.
11. A repeatable troubleshooting playbook
When a research service fails, change one variable at a time. First reproduce the issue with Clash disabled or with the institution’s documented baseline. Next enable Clash with system proxy mode and a simple policy. Then inspect DNS, selected rules, and the exact connection outcome. If the browser succeeds but the desktop application fails, compare processes and proxy inheritance. If only attachments fail, isolate storage endpoints. If only login fails, map the redirect and identity-provider domains.
- Record the symptom: distinguish a DNS error, timeout, TLS error, login loop, partial page, failed upload, or stale compilation result.
- Capture the destination: use Clash logs and browser tools to identify the hostname and the rule that matched it.
- Check the route: confirm whether the connection is
DIRECT, sent to the intended group, or accidentally caught byMATCH. - Test the application boundary: verify system proxy support, environment variables, helper processes, and TUN behavior.
- Revert broad experiments: remove temporary suffix rules after testing and replace them with the narrowest confirmed rule.
Keep a known-good backup of your profile before making changes. If you use remote rule providers, make local overrides explicit and check their precedence. A remote list may update between two tests, creating the illusion that your local edit caused a change. Version-control your configuration when possible, and write a short reason beside every non-obvious exception.
12. Frequently asked questions
Should Scholar and Overleaf share one proxy group?
They can, but they do not have to. A shared RESEARCH group is simple when both services work well through the same stable exit. Separate groups are better when Scholar searches need one region while Overleaf collaboration, Git hosting, or large uploads perform better through another. Start with one group, measure real tasks, and split only when the evidence justifies the added configuration.
Why does Zotero fail when Scholar works?
Scholar is usually tested in a browser, while Zotero is a separate desktop process with its own sync, storage, translator, and authentication connections. Confirm that Zotero inherits the intended proxy, then force a sync and inspect every endpoint it contacts. A working browser rule does not automatically cover WebDAV or an institution-managed storage service.
Should TUN mode be enabled immediately?
No. Begin with system proxy mode because it is easier to understand and less disruptive to campus networking. Enable TUN when logs show that a required application bypasses the system proxy or ignores environment variables. After enabling it, test private networks, institutional VPNs, SSH, printers, virtual machines, and the research workflow itself.
13. Keep the configuration measurable and maintainable
A durable research profile is not the profile with the most domains. It is the one whose rules have a clear reason, whose exceptions protect institutional traffic, and whose policy groups match the way researchers actually work. Review the endpoint notebook after major Zotero, browser, Overleaf, or university identity-provider updates. Retest a complete workflow instead of checking only a homepage, and remove rules that no longer appear in logs.
Compared with one-click VPN clients that apply a single route to every application, a carefully maintained Clash profile requires more observation but gives you a practical advantage: Google Scholar can use a reliable research path, Zotero can be tested as its own application, Overleaf can receive a stable collaboration route, and campus services can remain direct or institutionally managed. If you want that level of control across your workstation, you can choose a Clash client for your platform and begin with a measured research profile →
Ready to get started?
Build a cleaner research workflow with precise routing for Scholar, Zotero, Overleaf, and the campus services you rely on every day.