1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Start Codex in an empty project folder

export RELAY_API_KEY="<your relay key>"
mkdir sorter && cd sorter
codex

4. The prompt

Write sort_folder.py. Take a folder path as an argument. Move files
(not subfolders) into Images, Documents, Archives, Installers,
Audio, Video and Other based on extension. Never overwrite: if a
name exists, add _1, _2. Default to a dry run that only prints the
moves. Only move files when I pass --apply. Write every move to
moves.log. Use only the standard library.

5. Dry run first, then apply

python sort_folder.py ~/Downloads
python sort_folder.py ~/Downloads --apply

Read the dry-run list before you add --apply. If something lands in Other, ask Codex to add that extension to a group.

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Start Codex next to your files

export RELAY_API_KEY="<your relay key>"
cd ~/exports
codex

4. The prompt

Write merge_csvs.py. Read every .csv in a folder given as an
argument. Match columns by header name, not position. Keep the
union of all columns and leave cells empty where a file lacks one.
Add a source_file column. Write merged.csv and print how many rows
came from each file. Do not modify the input files. Use only the
standard library.

5. Run and check

python merge_csvs.py ~/exports
head -5 merged.csv

Add up the per-file row counts it prints and compare with the row count of merged.csv. If a file shows garbled text, ask Codex to detect its encoding.

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Start Codex in a new folder

export RELAY_API_KEY="<your relay key>"
mkdir dupes && cd dupes
codex

4. The prompt

Write find_dupes.py. Take a folder path as an argument. Group files
by size first, then by SHA-256 hash, so only files with identical
content count as duplicates. Print each group with full paths and
sizes. By default delete nothing. Add a --move-to option that moves
all but the oldest file in each group into a folder I choose. Never
delete files outright.

5. Run and check

python find_dupes.py ~/Downloads
python find_dupes.py ~/Downloads --move-to ~/dupes-review

Look through ~/dupes-review before you empty it. Run the first command on a small test folder before your real one.

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Set up the folder

export RELAY_API_KEY="<your relay key>"
mkdir thumbs && cd thumbs
cp -r ~/Pictures/trip ./photos
python -m venv .venv && source .venv/bin/activate
pip install pillow
codex

4. The prompt

Write make_thumbs.py. For every .jpg, .jpeg and .png in ./photos,
save a copy into ./thumbs that fits inside 400x400 and keeps the
aspect ratio. Respect the EXIF rotation so phone photos are not
sideways. Keep the same file name. Skip files that already exist
in ./thumbs. Never modify anything in ./photos. Print a count at
the end.

5. Run and check

python make_thumbs.py
ls thumbs | wc -l
ls photos | wc -l

Open a few thumbs to make sure none are rotated. Want a different size? Ask Codex to add a --size option.

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Set up the folder

export RELAY_API_KEY="<your relay key>"
mkdir excel-clean && cd excel-clean
cp ~/Downloads/report.xlsx .
python -m venv .venv && source .venv/bin/activate
pip install pandas openpyxl
codex

4. The prompt

Read report.xlsx (first sheet) with pandas. First print the first
15 rows and the column names so we can see the mess.
Then write clean.py that: finds the real header row, drops fully
empty rows and columns, strips spaces from text, parses the Date
column into YYYY-MM-DD, and writes clean.csv.
Never overwrite report.xlsx.

5. Run and check

python clean.py
head clean.csv

Compare the row count with the sheet. Next month, drop the new file in and run python clean.py again.

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
mkdir tiny-api && cd tiny-api
git init
codex

4. The prompt

Create a FastAPI app in main.py with one endpoint:
POST /slug takes {"text": "..."} and returns {"slug": "..."}
(lowercase, spaces to dashes, strip other symbols).
Return 422 if text is empty.
Add test_main.py using fastapi.testclient and pytest.
Add requirements.txt with fastapi, uvicorn, httpx, pytest.

5. Run it

pip install -r requirements.txt
pytest
uvicorn main:app --reload

Open http://127.0.0.1:8000/docs and try the endpoint from the page.

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
cd my-script
git add -A && git commit -m "before retries"
codex

4. The prompt

Add a fetch_with_retry(url, **kwargs) helper in http_utils.py using requests.
Retry up to 4 times on timeouts, connection errors, 429 and 5xx.
Wait 1, 2, 4, 8 seconds between tries. Respect Retry-After if present.
Do not retry other 4xx errors.
Replace the requests.get calls in main.py with it. No other changes.

5. Check it

git diff
python main.py

• There is a timeout on every request

• A 404 fails right away instead of waiting

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
cd my-script
git add -A && git commit -m "before sqlite"
codex

4. The prompt

Add a db.py using sqlite3 from the standard library.
Create a table results(id, created_at, value) if it does not exist.
Add save_result(value) and latest(n=10).
Change main.py to call save_result instead of writing results.txt.
Use parameterized queries. No other changes.

5. Check it

python main.py
sqlite3 data.db "select * from results order by id desc limit 5;"

• Queries use ? placeholders, not f-strings

• data.db is in .gitignore

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
cd my-repo
codex

4. The prompt

Create rate_limit.py with a decorator @rate_limit(calls=5, per_seconds=1).
If the limit is hit, sleep until the next window.
Use only the standard library. Add a short docstring. Nothing else.

5. Check before you commit

• No network calls inside the decorator itself

• No keys written into the file

• Import it and call a dummy function in a loop to confirm it paces

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
cd my-repo
codex

4. The prompt

Create .github/workflows/test.yml.
Run on push and pull_request.
ubuntu-latest, Python 3.12.
pip install -r requirements.txt, then run pytest.
Nothing else.

5. Check before you push

• File path is exactly .github/workflows/test.yml

• No keys or tokens written into the YAML

• Then commit, push, and open the Actions tab on GitHub

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• Model not found: copy the exact model ID, same case and hyphens

Relay cards: jliushi.github.io/ai-relay

1. Use a relay that supports Codex

On the AI Relay list, AgentRouter is marked for Claude Code and Codex: $100 free sign-up credit at 1x, GitHub account created before 2026. artbloom is Claude Code only — skip it for this.

2. ~/.codex/config.toml

model = "<model ID copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
cd my-app
codex

4. The prompt

Write a Dockerfile for this Python app.
Use python:3.12-slim.
Copy requirements.txt, install deps, then copy the app.
EXPOSE 8000. CMD runs python app.py.
No multi-stage build. Keep it short.

5. Review before you build

• No secrets baked into the image

• COPY paths match your real files

• Then: docker build -t my-app .

If Codex errors

• 404: check base_url and /v1

• 401 / 403: wrong key or credit used up

• 429: wait or switch relays

Relay cards: jliushi.github.io/ai-relay

1. Get a key

Pick a relay from the list. artbloom has no special sign-up rules. Create a token in the dashboard and copy the Base URL.

2. Set two variables

export ANTHROPIC_BASE_URL="https://<relay address>"
export ANTHROPIC_AUTH_TOKEN="<your relay key>"

3. Save your work, then start

cd my-script
git init 2>/dev/null; git add -A && git commit -m "before fix"
claude

4. Paste the full error

Running python app.py gives this error:
<paste the whole traceback>
Tell me the cause in two sentences, then fix only that.

Paste the whole traceback, not just the last line. The file name and line number matter.

5. Run it again

git diff
python app.py

New error? Paste that one. Same error? Say so and add what you tried.

Keep the credit

• One error per message.

• Start a new session when the chat gets long.

• Don't paste real passwords, keys or customer data. The relay sees everything you send.

Relays with free sign-up credit: jliushi.github.io/ai-relay

1. Pick a relay that works with Codex

On the AI Relay list, AgentRouter's card says Claude Code and Codex both work: $100 free sign-up credit at 1x, GitHub account from before 2026. Create a token after you log in.

2. ~/.codex/config.toml

model = "<model name copied from the relay's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
mkdir backup-tool && cd backup-tool
codex

4. The prompt

Write backup.sh. Take a folder path as the first argument,
zip it to ~/backups/<foldername>-YYYY-MM-DD.zip,
create ~/backups if missing, and print the output path.
Exit with an error message if the folder doesn't exist.

5. Test it on something harmless

chmod +x backup.sh
mkdir -p /tmp/demo && echo hi > /tmp/demo/a.txt
./backup.sh /tmp/demo

Read the script before you point it at real folders.

Next small asks

• "Keep only the 7 newest backups."

• "Add a --dry-run flag."

If Codex errors

• 404: base_url is missing /v1

• 401: key not exported, incomplete, or credit used up

• 429: rate limit. Wait or switch relays.

Card details and config notes: jliushi.github.io/ai-relay

1. Register and get a key

AgentRouter signs up with GitHub, and the account must be from before 2026. Balance shows $0? Log out and log back in. Then create a token.

2. ~/.codex/config.toml

model = "<model name copied from AgentRouter's list>"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

Copy the model string exactly. Don't change case or hyphens.

3. Export the key and start

export RELAY_API_KEY="<your relay key>"
mkdir csv2json && cd csv2json
codex

4. The prompt

Write csv2json.py. Read a CSV path from the command line,
use the first row as keys, print JSON to stdout.
Standard library only. Add a sample data.csv with 3 rows.

5. Try it

python3 csv2json.py data.csv

Next small ask: "Add an -o option to write to a file." One feature per prompt keeps the credit going further.

If it errors

• 404: check the /v1 at the end of base_url

• 401: key incomplete or credit used up

• 429: the site is rate limiting. Wait or switch sites.

Full card and setup notes: jliushi.github.io/ai-relay

1. Open a WSL terminal

Ubuntu (or your distro) from Windows Terminal — not PowerShell alone.

2. Set the two variables in Linux

export ANTHROPIC_BASE_URL="https://<relay address>"
export ANTHROPIC_AUTH_TOKEN="<your relay key>"
# optional if the model list fails to load:
export ANTHROPIC_MODEL="claude-opus-5"

3. Make them stick

nano ~/.bashrc

Paste the same exports at the bottom, save, then:

source ~/.bashrc

4. New WSL window, then test

echo $ANTHROPIC_BASE_URL
claude

Common mix-up

• Windows setx only helps native Windows apps. WSL Claude Code ignores it.

• Empty echo in WSL means the export never landed in the Linux shell.

Relay list and setup notes: jliushi.github.io/ai-relay

Before you start

• Claude Code is pointed at a relay with ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN.

• You opened a new terminal after setting them.

• This is a personal project. A relay sees everything you send.

1. Start a folder

mkdir todo-cli && cd todo-cli
claude

2. First prompt

Create todo.py: a command-line to-do list.
Commands: add "text", list, done N.
Save tasks to tasks.json. Standard library only.

3. Run it

python3 todo.py add "buy milk"
python3 todo.py list
python3 todo.py done 1

4. Two follow-up prompts

• "Add a clear command that removes finished tasks."

• "Write a README with usage examples."

Keep credit use low

• One small task per prompt.

• Run /clear when you switch to a new task.

• Don't paste big files you don't need.

If a prompt errors

• 401: key or credit.

• 404: Base URL.

• 429: the relay is rate limiting. Switch sites.

Pick a relay with free sign-up credit: jliushi.github.io/ai-relay

What the list says

• Daily free quota
• Runs on edge nodes
• Meant for apps you wire up yourself

What it is not

• Not a Claude Code relay. It won't replace ANTHROPIC_BASE_URL + a relay key.
• Not the same models as Opus on the relay list.
• Not unlimited. Daily free quota means limits can reset and change.

When to use it

1. Relay credit is gone and you need a spare for a non-Claude tool or a small app.
2. You already have a Cloudflare account and can check the current free limits there.
3. You keep company code and secrets off any third-party path.

For Claude Code or Codex, pick a relay first. Workers AI is the backup lane.

Relays and free-tier backups: jliushi.github.io/ai-relay

1. Save the key to Keychain

security add-generic-password -a "$USER" -s relay-key -w "<your relay key>"

2. Load it in ~/.zshrc

export ANTHROPIC_BASE_URL="https://<relay address>"
export ANTHROPIC_AUTH_TOKEN="$(security find-generic-password -a "$USER" -s relay-key -w)"

3. Open a new terminal and test

echo ${ANTHROPIC_AUTH_TOKEN:0:6}
claude

Say hello in Claude Code. If it answers, you're done.

Changing or removing the key

• New key: security delete-generic-password -s relay-key, then run step 1 again.

• Codex: use the same trick for RELAY_API_KEY, since config.toml only names the variable.

If it breaks

• Empty output in step 3: the service name doesn't match. Use relay-key in both commands.

• 401: the key you saved is incomplete. Copy it again from the relay dashboard.

Relay list and setup guide: jliushi.github.io/ai-relay

What env_key actually does

It only names the environment variable. It does not store the key. If that variable is empty in your shell, Codex has nothing to send.

Fix

export RELAY_API_KEY="<your relay key>"
codex

To keep it after you close the terminal, add the same export to your shell config (zshrc, bashrc, or PowerShell profile), then open a new terminal.

Minimal config.toml reminder

model = "gpt-5.6"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

Copy the model name from the relay's own list. Keep /v1 on the Codex base_url. Don't paste the key into config.toml.

Quick checks

• echo $RELAY_API_KEY — should print your key, not blank.
• 401: key missing, truncated, or credit gone.
• 404: base_url /v1 problem, not the env var.

Relay list and Codex notes: jliushi.github.io/ai-relay

1. Pick a relay that supports Codex

On the tested list, AgentRouter is marked for both Claude Code and Codex. Relays marked Claude Code only won't work with Codex. Sign up, then create a key on the Tokens (令牌) page and note the base URL.

2. Edit config.toml

mkdir -p ~/.codex
nano ~/.codex/config.toml

Paste this and fill in the relay's address:

model = "gpt-5.6"
model_provider = "relay"

[model_providers.relay]
name = "relay"
base_url = "https://<relay address>/v1"
env_key = "RELAY_API_KEY"
wire_api = "chat"

Use a model name from the relay's own list, copied exactly. Keep /v1 at the end of base_url for Codex.

3. Put the key in ~/.zshrc

echo 'export RELAY_API_KEY="<your key>"' >> ~/.zshrc
source ~/.zshrc

The key goes in an environment variable, not in config.toml.

4. Run it

codex

If it fails

• 404: check the base URL and /v1.
• 401: check the key and your remaining credit.
• Errors right away on every request: keep wire_api = "chat".

Relay list and the full Codex setup notes: jliushi.github.io/ai-relay