ai-coding-welfare: a script-generated directory of free Claude Code and Codex relay stations

What ai-coding-welfare actually collects

The repository is a directory, not a proxy. It does not sit between your client and a model provider, and it does not issue API keys. What it holds is a table of ten stations that hand out free or discounted access to Claude and GPT-family models, plus the metadata needed to compare them: signup bonus, referral bonus, daily check-in amount, supported wire protocols, and a link to register. The README describes the target audience indirectly but clearly. Someone who wants to run Claude Code or Codex CLI without a paid Anthropic or OpenAI account, and who is willing to register on a station they have never heard of in exchange for a few dollars of daily quota.

The units are inconsistent on purpose, because the stations are. Eight entries are priced in dollars; Matrix hands out 600 points and NOFX 15 points, and the README states plainly that these are in-site points with no published conversion to dollars, so they are excluded from the $817 first-day total. That total covers the eight dollar-denominated stations that still accept new users. The README also flags that four stations have gone more than 48 hours without a successful interface fetch, meaning their detail rows are the last snapshot rather than live data.

How the data pipeline works: scrape, store, rebuild

The repository layout tells most of the story. There is a data/ directory, a scripts/ directory, and a package.json whose scripts read like a pipeline: refresh, history, build, check, notify, test. The README states that model lists, prices and online status are generated by scripts that fetch the stations' public interfaces, and that the last update was 2026-09-14 12:51 UTC. The npm script named all chains them: npm test, then refresh, then history, then build, then check.

So the flow is: refresh.mjs pulls current model and pricing data from each station's public endpoint; history.mjs appends a snapshot, with a --backfill flag exposed as npm run backfill for filling in past data; build.mjs regenerates the README and the landing page from what is in data/; check.mjs validates the result; notify.mjs handles notifications. The releases are named after that data, not after code. data-20260908-1122 is labelled "CheapCodex 恢复在线" and data-20260908-0439 is labelled "CheapCodex 探测不到了：注册页与公开接口都没响应", which is the same station going down and coming back within about seven hours. The changelog and the Atom feed at feed.xml are outputs of the same cycle.

The consequence is that the repository's freshness depends entirely on the stations exposing a scrapeable public interface. The AgentRouter entry says its interface has gone 311 hours without new data, and the snapshot shown is from 2026-09-01 14:01 UTC. A station can be perfectly usable while its metadata is frozen, and the reverse is also true.

Installing and configuring Claude Code through a station

There is nothing to install for the directory itself unless you want to regenerate it. For that you need Node 18 or newer, per the engines field, and then the pipeline scripts.

# requires Node >= 18
git clone https://github.com/panxunying/ai-coding-welfare.git
cd ai-coding-welfare
npm run refresh && npm run history && npm run build && npm run check

The more common path is using the directory to configure a client. The README's three-step version is: register through the table's link (it warns that quota is issued against the referral link, so a bare URL will not credit you), create a key under the station's tokens or API Keys page, then either run the quickstart script or copy the environment variables by hand. The quickstart scripts are interactive and write the Claude Code variables for you.

# macOS / Linux
bash scripts/quickstart.sh

# Windows PowerShell
powershell -ExecutionPolicy Bypass -File scripts/quickstart.ps1

If you prefer to set them yourself, the AgentRouter example in the README uses an Anthropic-compatible base URL without a /v1 suffix, and installs the CLI in the same block.

export ANTHROPIC_BASE_URL=https://agentrouter.org
export ANTHROPIC_AUTH_TOKEN=你在站点后台创建的 Key
export ANTHROPIC_MODEL=claude-opus-5
npm install -g @anthropic-ai/claude-code@latest && claude

For Codex CLI the README writes a provider block into ~/.codex/config.toml instead, with base_url carrying the /v1 suffix and env_key naming the variable that holds the key.

model = "gpt-5.6-sol"
model_provider = "agentrouter"

[model_providers.agentrouter]
name = "AgentRouter"
base_url = "https://agentrouter.org/v1"
env_key = "AGENTROUTER_API_KEY"
wire_api = "chat"

Generic OpenAI-compatible clients need only two fields: Base URL set to https://agentrouter.org/v1 and the API key from the station dashboard. The README also gives a curl one-liner against /v1/chat/completions as a connectivity self-test, which is the fastest way to tell whether a station is actually serving before you blame your client configuration.

What breaks, and why the online badge is weaker than it looks

The failure modes here are operational, not architectural. Stations disappear. The CheapCodex release pair is the clearest evidence in the repository: a probe failure at 2026-09-08T04:39:32Z and a recovery at 2026-09-08T11:22:04Z. RawChat shows as 异常 in the table while the other nine show as online. AgentRouter, the station the README recommends first, has not yielded interface data for 311 hours. Four stations are running on stale snapshots.

The README is upfront that online status is judged by whether the registration page is actually reachable, not by whether the API endpoint answers. Those are different checks, and a station can pass the first while failing the second. The quota numbers have the same character. First-day totals assume the signup bonus, the referral bonus and the same-day check-in all land, and the README notes that for daily-reset stations the pool refills the next day but does not accumulate. So the $817 headline is a one-day figure, not a balance you keep.

This is the wrong tool if your prompts cannot be sent to an unvetted third party, if you need a base URL that will still resolve next quarter, or if you are wiring an automated pipeline where a silent endpoint change breaks the build. It is a tool for trying things out.

Compared with running LiteLLM or a self-hosted gateway

The obvious alternative is a self-hosted gateway such as LiteLLM in front of provider API keys you control. The difference in approach is where the trust and the cost sit. LiteLLM is software you run: you supply the upstream credentials, you own the routing, retries and logging, and the cost is whatever the providers charge plus the machine you run it on. ai-coding-welfare is a directory of upstreams you do not control, and its value is the comparison data and the setup scripts, not any code on the request path.

That means the two solve different halves of the problem. A gateway gives you one stable base URL and consistent behaviour across models; this repository gives you ten candidate base URLs, some of which are free, none of which promise stability. If you already pay for Anthropic or OpenAI access, the gateway is the better fit and this repository has nothing to offer you. If the entire point is avoiding that payment, no gateway helps, because the gateway still needs a funded key behind it.

A second alternative is simply reading the stations' own documentation and configuring one client by hand. That works, and it is what the README's manual path describes. The reason to use the repository instead is the refresh pipeline: model lists and prices change, and the README and landing page are regenerated from data/ rather than edited by hand, so the comparison table does not drift silently. That is a real difference, but it is a maintenance convenience, not a capability.

Maintenance cost, the MIT licence, and what you are agreeing to

The repository is MIT-licensed, which covers the scripts, the generated README and the landing page. It does not cover the stations, their terms of service, or the models behind them. Nothing in the licence grants you access to any API, and the README's own framing is that registration and quota are governed by each station. If you fork this and point it at your own endpoints, MIT gives you the usual permissions; if you assume the licence extends to the relay services, you have misread it. This is not legal advice, and the terms that matter are the ones you accept when you register.

The last push was on 2026-09-14, so the pipeline is current as of that date. The maintenance burden falls in two places. Upstream, someone has to keep the scrapers working as station interfaces change, and the release names show that probe failures are treated as events worth publishing. Downstream, if you clone it, you inherit that burden: a station that changes its public endpoint shape will break refresh.mjs, and the check step exists to catch a broken build before it ships. Running npm run all locally is the honest way to find out whether the data you are reading is still being fetched or is a frozen snapshot.

Editorial conclusion

Adopt this if you want to try Claude Code or Codex CLI without paying, accept that the endpoints behind it are third-party relays with no uptime guarantee, and are willing to re-run scripts/refresh.mjs when a station changes. Do not adopt it if your prompts cannot leave your machine, or if you need a stable base URL in CI. Verify two things first: whether the station you pick is one of the four whose interface data is more than 48 hours stale, and whether the quota you were promised actually appears in the dashboard after you log out and back in, which the README says is required for AgentRouter credits to display.

Frequently asked questions

Official sources

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