The day-one total counts referral credit, and one station's row is two days stale

The README, the landing page and the releases are all script output

Nothing on the page is maintained by hand. The manifest declares a private package named `ai-coding-welfare` at version 1.0.0, requiring Node.js 18 or newer, and every script is one Node file. `refresh` runs scripts/refresh.mjs, `history` runs scripts/history.mjs and `backfill` runs the same file with `--backfill`, `build` runs scripts/build.mjs, `check` runs scripts/check.mjs, `notify` runs scripts/notify.mjs, and `test` runs scripts/test.mjs. The aggregate command runs them in order: test, refresh, history, build, check.

The published releases are data snapshots rather than code versions. The three newest tags are `data-20261002-1238`, `data-20261001-1319` and `data-20261001-0633`, each carrying a Chinese sentence describing what changed rather than a version number. Two of those three sentences are about the same station: one records that FlushAPI is back online, and the previous day's records that its registration page and public endpoints both stopped responding.

That naming is the clearest statement of what the repository is. A tag per refresh means the history of the page is the history of what the stations were doing, and the code itself has no released version beyond the 1.0.0 in the manifest. A private marker on the package keeps it off npm, which is right, since the artefact here is a page and a data directory.

The day-one total is a formula that depends on arriving through the links

The headline column has a stated definition rather than a marketing gloss: day-one credit equals the registration base credit plus the credit attached to this page's invitation link plus whatever check-in credit is claimable that day, with sites that reset a daily pool counted as one day's pool. Summing the eight sites that publish a dollar figure and still accept new users gives roughly $1199.5 for day one.

Three sites pay in their own units instead and are excluded from that total because no public conversion exists: DoCode issues 300 of its in-site currency, Matrix 600 points, and NOFX 15 points. So the total is a floor over the convertible sites, not a measure of everything on offer.

The formula is also why the page tells you not to arrive through a bare link. Every registration link on it carries a referral parameter belonging to the page owner, and the credit attributed to that parameter is one of the three terms in the sum. Quoting one station's arithmetic: registration $75, this page's invitation $25, first sign-in $20, giving $120 on day one and $20 a day thereafter.

The invitation-code column records the other half of this. One site publishes a code that has to be typed into the registration form by hand; leave it out and you keep only the base credit, which the page says cannot be claimed afterwards. Rows marked with a dash mean no such code is registered.

Online status comes from the registration page, not from the API

Two different probes produce two different columns, and the page is explicit about which is which. The availability column is judged by whether a station's registration page is actually reachable. The quota and price figures are scraped from each station's public endpoints. When a station has not answered for more than 48 hours, its detail row keeps the last successful snapshot and says so.

That is why one row in a table of thirteen can be two days behind the others while the status column still reads green for all thirteen. A registration page that loads does not prove the API behind it is answering, and a stale quota snapshot does not mean the station is down. Anyone automating against this table has to carry both signals and know which one they are reading.

The scan output is also visible per station, at least where the page shows it: ArtBloom's entry records an interface latency of 790 ms. Elsewhere the models column is honest about its own gaps, reading that login is required to see the list for most rows. The two exceptions give exact figures with a date attached: Opus 5.5 at an input price of $2 per million tokens and an output price of $10 per million, verified against the public model page on 2026-09-28, with the model id given as `claude-opus-5-5`.

Most stations hide their quota behind a login or a bot

The protocol column is where the practical work is, and it is not uniform. Some stations speak OpenAI only. Several speak Anthropic plus OpenAI. Two are marked as OpenAI compatible without naming a vendor. Mirasim is reached through its own client or CLI rather than a documented HTTP surface. Two quotas are visible only inside the station's console, and one only inside a Telegram bot, which is why a machine-readable table cannot carry every figure.

The model column follows the same pattern. A few rows list names outright, such as a trio of GPT variants on one station and a GPT plus Fable pairing on another, and one says four are visible. Most say login is required. So the table tells you which station speaks which protocol but not, in most rows, what you can actually call once you are signed in.

Prices are given for one model on one station, and the page treats that as a reference rather than a rate card. One note explains a unit that looks like a monthly allowance and is not: a $1 monthly plan on one station is estimated per five-hour window on a shared quota converted by model, the per-model counts cannot be added together, and paid plans are excluded from the free-credit total. That pricing was checked on 2026-09-23, ten days before the data timestamp printed on the page.

The quickstart writes your editor's own config file

The three-step path is: open a registration link and sign in the way the station supports, create a key in its token or API keys page, then run the script or copy the environment variables by hand. The script needs Node.js on macOS or Linux:

# 交互式写好 Claude Code 的环境变量（macOS / Linux，需要 Node.js）
git clone --depth 1 https://github.com/panxunying/ai-coding-welfare.git && cd ai-coding-welfare
bash scripts/quickstart.sh

Windows gets a separate script, and the note attached to it is the kind of detail that saves an afternoon: PowerShell 5.1 does not accept `&&`, so the two commands are separated with a semicolon instead.

# Windows PowerShell（5.1 不认 &&，两条命令用 ; 隔开）
git clone --depth 1 https://github.com/panxunying/ai-coding-welfare.git; cd ai-coding-welfare
powershell -ExecutionPolicy Bypass -File scripts/quickstart.ps1

What it writes depends on the client. For Codex CLI it edits `~/.codex/config.toml`, adding a provider block with a base URL, an environment key name and `wire_api = "chat"`. For a generic client such as the OpenAI SDK, Cherry Studio or Cursor, the requirement is two fields: a base URL ending in `/v1` and the key you created. The script's job is to write those files, so read them before you run it.

Six languages of README over four generated sub-pages

The repository root carries README.md plus five translations, README.en.md, README.hi.md, README.pt-BR.md, README.ja.md and README.de.md, and an I18N.md whose purpose the root does not spell out. The landing page at the GitHub Pages address mirrors that reach with its own sub-pages: a per-request versus per-volume converted comparison at `/compare/`, an availability history at `/status/`, a change log at `/changelog/`, and an Atom feed at `/feed.xml`.

The availability history and the feed are the parts with a durable use. A station that is green today may have been unreachable yesterday, and the release tags record that in the same way the file history does. Reading the change log tells you whether a station's outage was momentary or whether it has been quiet for a while.

Contributions run through the issue tracker, and the page is direct about the arrangement: submissions are public, the issue area is itself a promotion channel, and the ones found to be genuinely good get written into the main text. That is a fair trade for a directory, and it also means the table is partly a function of what operators submit about themselves.

The data timestamp on the page runs ahead of the last commit

One number does not line up. The page prints its own last-update time as `2026-10-02 22:16 UTC` for the scraped models, prices and status. The repository's last recorded push is 2026-10-01T13:19:32Z, and the newest data tag is `data-20261002-1238`, which is 12:38 UTC on the second, some nine hours before the time the page claims for its data.

A generated page will always be able to run ahead of the branch it was generated from, and the release tags being newer than the recorded push is a sign the tags are cut from somewhere other than the branch head. Neither is harmful on its own. What it does mean is that the timestamp on the page is a claim by the generator rather than something a reader can confirm from the commit history, so it is worth treating the git tags as the version of record.

Licensing here is the one unambiguous thing. The manifest declares MIT and a LICENSE file sits at the root, and the repository is not archived. The code that scrapes, ranks and rewrites the directory is permissively licensed; the stations it points at are separate services with their own terms.

Editorial conclusion

This page is best read as a monitoring board rather than a recommendation. It states its own limits in unusually plain language: one row can be two days old, most stations hide their quota behind a login, and the headline figure counts credit that only arrives if you enter through the page's own referral links, so the number is a measure of the arrangement the page sets up rather than a measure of what any station is worth on its own. Three checks before you point a real project at any of these endpoints. Confirm the terms of both the model provider and the station, because a relay in the middle is not the provider. Test latency and model identity against the station's own model page rather than trusting the table, since the table's figures are pulled from public endpoints that may be stale or may not require a login. And read the entry for the station you intend to use: the protocol column mixes OpenAI, Anthropic plus OpenAI, OpenAI-compatible, a vendor client and a Telegram bot, so the integration work differs per row. The one quantity the page does not offer is what any of it costs after the free credit runs out.

Frequently asked questions

Official sources

1. License: MIT
2. panxunying/ai-coding-welfare on GitHub
3. Project website
4. README
5. Releases