freeSub: a scored proxy node subscription in three client formats

A repository of configuration files, not a generator

The description is short: a free clash subscription, updated at any time, with a repeated instruction not to fork it. The homepage is a Telegram group link, and there are no GitHub releases.

The top level entries tell you what this project is. There is a README.md, a FUNDING directory, a .github directory, and exactly three payload files: clash.yaml, singBox.json and v2ray. There is no source directory, no build script, no test suite, no package manifest and no code of any kind in the listing.

That matters for how you evaluate it. You are not adopting a tool. You are subscribing to a set of configuration files that somebody else regenerates and publishes, and your client consumes them over the network at an interval you set. The last push was on 2026-10-18, and the README carries a timestamp row showing an update at 2026-10-18 22:00:01 with 500 available nodes.

The consequence is that you can only verify the output and the conventions. The scoring model, the service access detection and the node generation all happen somewhere else, on a machine you do not control, and the repository does not document that process beyond describing the results.

What the repository is unusually good at is explaining the format of what it publishes. The node naming scheme, the risk grade table, the scoring dimensions and the scenario groups are all documented in detail, so a user can read a node name and understand what it means before selecting it. That is a better state of affairs than most subscription lists, where the names are opaque strings.

Three formats, one subscription step

The delivery mechanism is the same for every client, and it is worth understanding because the accelerated and raw addresses behave differently.

For Clash and Mihomo the recommended address runs through a GitHub proxy, and the README notes that an accelerated link is recommended because a direct connection is more stable from mainland China. The raw GitHub address is offered as a backup.

The addresses for all three formats are published as plain URLs in the README:

https://gh-proxy.com/raw.githubusercontent.com/Ruk1ng001/freeSub/main/clash.yaml
https://gh-proxy.com/raw.githubusercontent.com/Ruk1ng001/freeSub/main/v2ray
https://gh-proxy.com/raw.githubusercontent.com/Ruk1ng001/freeSub/main/singBox.json

The pattern repeats for the other two formats, and for the V2Ray and Shadowrocket entry the same two addresses are given, the accelerated one first and the raw one in a collapsed backup section. SingBox follows with its own pair. The three client formats are therefore covered by six URLs, all of them pointing at the same repository with a different path.

The usage steps are three, and they are the same in every client. Copy the subscription address for your client, open the proxy client and import the subscription, then select a scenario group and turn the proxy on.

What the user does not control is the content. The file is replaced wholesale on each refresh, so a subscription that worked this morning may present a different node set tonight. There is no version history in the repository to diff against, and no tagged release to pin to, which is a real operational consideration for anyone who wants to know what changed.

Node names carry the metadata the client UI cannot show

The most substantive part of the documentation is the node naming scheme, because it encodes a per-node assessment into a string you can read in a client list.

Each name is composed of segments, and the README draws the structure with a diagram. The first segment is the quality grade, one letter plus a sign. Then a risk level icon, then a country flag derived automatically from the GeoIP country code, then a bracketed group of capability icons limited to at most four, then the sponsor identifier, and finally the first eight characters of the node id.

The quality grade is the part to understand first, and the documentation is explicit that it combines the IP type, whether residential or datacenter or CDN, with a risk rating. The table gives five levels with a color for each. The green level, A+ or A, means a clean IP, mostly residential or low risk, with the highest success rate for AI and streaming. The yellow B is average, fine for daily use, though some risk-controlled sites may ask for verification. The orange C is higher risk, mostly datacenter IPs, and easily rejected by AI services. The red D is high risk, already widely flagged, and recommended only for emergencies. The white U is unknown, meaning the GeoIP data is missing.

The advice the documentation gives is short and practical: when the details are hard to read, look at the color, prefer green, then yellow.

The capability icons in brackets are described as services that were actually tested as reachable on that node, ordered by AI, then streaming, then community and tools, with at most four shown. The AI list covers the OpenAI API, the ChatGPT web interface, Claude, Gemini, Perplexity, Genspark and Microsoft Copilot. The streaming list covers Netflix, YouTube and YouTube Premium, Disney+, Spotify, TikTok, HBO Max and Prime Video. The community and tools list covers LinuxDO, Discord, Telegram, GitHub, Google and Wikipedia.

An important caveat follows. No brackets means the node did not pass any service access check, but basic connectivity is still working. And the documentation warns that a successful service check only means the connection works, not that an account will avoid risk control.

Composite score and scenario groups

The subscription ships a scoring model and a set of scenario groups, and the documentation's stated intent is that in most cases you select a group rather than testing nodes yourself.

The composite score runs from 0 to 100 and fuses several dimensions measured in real time. Latency and jitter, where lower is more stable and faster. Download speed. Availability, described as the historical success rate over recent checks, where higher is more reliable. And IP quality, covering residential versus datacenter and the risk grade.

The proxy group selects the current best node by that composite score and, according to the documentation, keeps eliminating nodes that are unstable over time based on availability.

The scenario groups sit on top of that. There is an AI group for ChatGPT, Claude and Gemini style services, a streaming group for Netflix, YouTube, Disney+ and Spotify, and a general browsing group that picks the best composite score for ordinary traffic. The selection advice is to prefer a scenario group or a region group and let the client choose, and to switch groups when something feels stuck.

There is also a region dimension, with flags for Hong Kong, the United States, Japan, Singapore, Taiwan, South Korea, Germany and the United Kingdom, and the guidance is to judge latency by distance and service access by the target service's regional restrictions.

This is a well thought out presentation of per-node data that a proxy client would otherwise have to compute itself. What it does not include is the measurement methodology, and the availability figure in particular depends on how often the checker runs, which is not stated.

The repository doubles as an advertisement

A large share of the README is not about the subscription at all. It promotes a separate commercial service, Dokng, described as an AI gateway or relay that fronts the major model providers, and the pitch is framed directly against the problems with free nodes.

The argument is a list of three. Free nodes are mostly datacenter IP addresses, so OpenAI and Anthropic refuse them or demand verification often. They are unstable, so a coding session drops mid-write and long Claude Code runs fail. And they require account maintenance, where registration, subscription, payment and bans each become a point of failure.

The relay is offered as the way around that, with a single base_url change and no new SDK. The documentation lists the entry points, one OpenAI compatible and one Anthropic compatible, both under the same domain, and a table of pricing multipliers by group, from a low tier opened at irregular times up to a higher multiplier for more channels, with the highest tier covering image and video generation.

The download tools section follows, describing a tool that takes over the configuration of the official desktop clients so you do not have to edit files by hand, and a mirrored copy of the official downloads so the client installers can be fetched without a cross-border download. The document states that both desktop clients are free and that they use your own key to reach the gateway.

This is worth naming plainly when you evaluate the repository. The subscription is the product, and the relay is the advertisement, and the two are connected because the free nodes have exactly the weakness the relay is positioned to fix. The subscription is still usable on its own, and the documentation's own warning that a successful check is not the same as avoiding risk control is the most honest line in the file. But you should read the promotional sections as marketing rather than as a feature list.

What to check before you point a client at this

Three things, and then the licence question.

First, check the timestamp and the node count in the header table before you import. The README shows 500 available nodes at the 2026-10-18 22:00:01 update. A file that is several days old is telling you the generation pipeline has stalled, and the node set you get will reflect that.

Second, understand the failure mode of the format you are using. A datacenter IP on a free node is exactly what the risk grading describes, and the grade C and D entries are documented as easily rejected by AI services and flagged in many places. If your use is development traffic that does not touch a service with aggressive risk control, the free list is a reasonable tool. If your use depends on those services, the documentation is telling you the list is the wrong instrument.

Third, remember there is no rollback. The files are replaced on each update, the repository has no GitHub releases, and there is no history of the previous configuration in the repository itself. If a client is configured against a subscription URL, the only stability is the client's own behavior, not the server's.

The licence is not stated. It is recorded as unknown, and there is no licence file in the top level listing, only a FUNDING directory and a README. A configuration file listing proxy endpoints is thin intellectual property, but the same applies to the node names and the scoring data, and if you intend to redistribute or build on the format, that question needs an answer before you commit.

The disclaimer in the README asks users to treat the project as for learning and exchanging information, not for any illegal activity, and states that all resources come from the internet and the repository only collects and organizes them, with an offer to remove infringing content through an issue.

There is one more thing to consider, and it is not the project's fault. Free proxy nodes are a moving target for everyone, not just for you, and the practical value of a subscription list of this kind fluctuates with the health of the upstream endpoints. Treat it as a convenience, not as infrastructure, and keep a paid or self-run option for the traffic that has to work.

Editorial conclusion

Adopt freeSub if you need a proxy subscription in a specific client format and the built in scoring and scenario groups are useful to you, and you are comfortable with nodes that are mostly datacenter addresses. Do not adopt it as a stable environment for AI workloads, because the documentation itself says free nodes are commonly refused by those services, and treat the promoted relay service as a separate commercial product rather than a feature of this repository. Verify first that the file you fetch is the one your client expects, that the timestamp in the header is recent, and that you have a fallback because the files are replaced wholesale on each update with no version history to roll back to.

Frequently asked questions

Official sources

1. Issues
2. README
3. Ruk1ng001/freeSub on GitHub