DEV Community: 江流石

Pin your relay config in ~/.claude/settings.json (no more export every time)

江流石 — Sun, 11 Oct 2026 01:12:11 +0000

Tired of export-ing env vars in every new terminal? Put your relay config in ~/.claude/settings.json under env, and Claude Code picks it up on launch:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://<station>",
    "ANTHROPIC_AUTH_TOKEN": "sk-****your-key****"
  }
}

Tips:

This file is local — don't commit it.
To juggle several stations, use ccswitch to keep each one on its own git branch and ccswitch switch between them.
If the model list won't load, add ANTHROPIC_MODEL with the exact id copied from the station.

Relay stations are free (signup + daily check-in credit) and plentiful — register a few. List with models and tested dates: AI relay station list

Signup links carry my referral param; strip ?aff= if you prefer.

Debugging a relay API in Claude Code / Codex: 401 vs 404 vs 429

江流石 — Sun, 11 Oct 2026 01:11:54 +0000

When you point Claude Code or Codex at a third-party relay and it errors, read the status code first — it tells you exactly what's wrong:

404 — wrong address. Usually the /v1 on base_url: Anthropic-compatible endpoints have no /v1 (Claude Code appends /v1/messages); OpenAI-compatible endpoints need /v1.
401 / 403 — auth failed. Key copied in full? Credit left? Model id in the station's list? Claude Code must use ANTHROPIC_AUTH_TOKEN (Bearer), not ANTHROPIC_API_KEY.
429 — the station is rate-limiting. Switch stations or wait; retrying just lengthens the queue.
400 / content blocked — upstream moderation. Retry in plain English/Chinese; if it keeps firing, switch stations.

Keep a couple of stations registered so a single one rate-limiting never blocks you. List with tested dates: AI relay station list

Signup links carry my referral param; strip ?aff= if you prefer.

Free Claude Code & Codex: official free tiers vs. relay stations

江流石 — Sat, 10 Oct 2026 13:33:54 +0000

Two ways to run Claude Code / Codex without paying per token:

1. Official free tiers — limited, but zero setup. Good for light use.

2. Relay stations — third-party gateways that mirror the API and hand out signup + daily check-in credit. More headroom, tiny bit of config.

Relay setup is two env vars (Claude Code, AgentRouter as example):

export ANTHROPIC_BASE_URL="https://agentrouter.org"
export ANTHROPIC_AUTH_TOKEN="sk-****your-key****"
claude

For Codex, point ~/.codex/config.toml at https://<station>/v1 with an OpenAI-compatible provider.

Notes: Anthropic endpoint has no /v1; OpenAI endpoint ends with /v1; keep the key in an env var. Copy the exact model id from the station.

These stations are free and plentiful — register a few and rotate. List with models + tested dates: AI relay station list

Signup links carry my referral param; strip ?aff= if you prefer.

Use GLM-5.3 for free through a relay API (OpenAI-compatible)

江流石 — Sat, 10 Oct 2026 10:21:17 +0000

GLM-5.3 is a solid, cheap model, and some free relay stations expose it with signup credit — no card needed. It's OpenAI-compatible, so any OpenAI-style client works.

Point your client at the relay:

export OPENAI_BASE_URL="https://<station>/v1"
export OPENAI_API_KEY="sk-****your-key****"
# model: glm-5.3

Or in Codex (~/.codex/config.toml):

model = "glm-5.3"
model_provider = "relay"
[model_providers.relay]
base_url = "https://<station>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

Notes:

OpenAI-compatible endpoint ends with /v1.
Copy the exact model id from the station's list.
Keep the key in an env var, not in the file.

Several free stations carry GLM-5.3 alongside Claude and Grok models — register a few and switch as needed. List with tested dates: AI relay station list

Signup links carry my referral param; strip ?aff= if you prefer.

Try Grok 4.7 free via a relay API

江流石 — Sat, 10 Oct 2026 10:12:38 +0000

Want to try Grok 4.7 without a paid plan? Some free relay stations expose it with signup + daily check-in credit. It's served over an OpenAI-compatible endpoint.

Minimal setup for any OpenAI-style client:

export OPENAI_BASE_URL="https://<station>/v1"
export OPENAI_API_KEY="sk-****your-key****"
# model: grok-4.7

Notes:

The endpoint ends with /v1 (OpenAI-compatible).
Use the exact model id from the station's model list.
Don't hardcode the key — use an env var.

These stations are free and usually carry several model families (Claude, GLM, Grok) behind one key, so you can register a couple and pick per task. Full list with models and tested dates: AI relay station list

The signup links carry my referral param — drop ?aff= if you'd rather not.

Vibe coding with Claude Code on free relay credits

江流石 — Sat, 10 Oct 2026 10:02:02 +0000

"Vibe coding" = you describe what you want, the agent writes and runs it. With Claude Code on a free relay station's credits, you can do it without paying for the official API.

Connect in two env vars (AgentRouter as the example):

export ANTHROPIC_BASE_URL="https://agentrouter.org"
export ANTHROPIC_AUTH_TOKEN="sk-****your-key****"
claude

Then just talk to it. Open an empty folder and say:

Build a todo app with Vite + React, local storage, and a dark mode.

Claude Code scaffolds the files, installs deps, runs it. You keep going in plain language — "add a tag filter", "make this green". Out of credit? Check in daily for more, or switch to another station.

These relay stations are free (signup + check-in credit), so grab a few. List with models and tested dates: AI relay station list

The signup links carry my referral param; strip ?aff= if you'd rather not.

Point Codex at a free relay API (config.toml walkthrough)

江流石 — Sat, 10 Oct 2026 10:01:46 +0000

Want to run Codex without paying per token? Point it at a third-party relay that mirrors the API and gives you signup credit. Here's a minimal ~/.codex/config.toml using AgentRouter.

model = "gpt-6-astra"
model_provider = "relay"
[model_providers.relay]
base_url = "https://agentrouter.org/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

Then export the key:

export RELAY_API_KEY="sk-****your-key****"

Notes:

OpenAI-compatible endpoint — base URL ends with /v1.
The key goes in an env var via env_key, not in the file, so the config is safe to back up.
Start with wire_api = "chat"; most relays expose chat completions, not the responses API.
Copy the exact model string from the station's model list.

Plenty of these free stations exist (signup + daily check-in credit), so register whichever you like. Full list with tested dates: AI relay station list

Signup links there carry my referral param — drop ?aff= if you prefer.

Run Claude Code on free credits via a relay API (AgentRouter)

江流石 — Sat, 10 Oct 2026 09:55:18 +0000

The official Claude API bills per token and wants a foreign card. A quicker way to try Claude Code for free: third-party relay stations that mirror the API and give you signup credit. Here's the AgentRouter setup.

Two environment variables:

export ANTHROPIC_BASE_URL="https://agentrouter.org"
export ANTHROPIC_AUTH_TOKEN="sk-****your-key****"
claude

Notes:

Anthropic endpoint — base URL has no /v1 (Claude Code appends /v1/messages itself).
Use ANTHROPIC_AUTH_TOKEN (sent as Authorization: Bearer), not ANTHROPIC_API_KEY.
Can't fetch the model list? Copy the exact model string from the station and set ANTHROPIC_MODEL.
To make it stick, put it in ~/.claude/settings.json under env. Don't commit your key.

Several of these stations exist — all free, with signup + daily check-in credit, so you can just register the ones you want. Full list (models, credits, tested dates): AI relay station list

The signup links on that page carry my referral param; strip ?aff= if you'd rather not.