Codex CLI supports custom model providers, which makes it a good fit for api.ish.chat. You add ISH as a provider in ~/.codex/config.toml, point it at the OpenAI-compatible /v1 endpoint, and tell Codex to use the Responses wire format.
This guide gives you the working config and the quick tests that prove text responses and tool calls are flowing correctly.
What you need
Before editing config:
- An ISH account.
- Purchased API balance. Codex/API traffic does not spend the free daily chat balance.
- An API key from api.ish.chat/keys.
- Codex CLI installed locally.
Keep the API key out of your repo. Put it in your shell environment or a private secret manager, not in a committed config file.
Add ISH as a Codex provider
Open:
~/.codex/config.toml
Add this config:
model = "claude-sonnet-4.6"
model_provider = "ish"
model_context_window = 200000
model_auto_compact_token_limit = 120000
[model_providers.ish]
name = "ISH API"
base_url = "https://api.ish.chat/v1"
wire_api = "responses"
env_key = "ISH_API_KEY"
The important pieces are:
base_url = "https://api.ish.chat/v1"wire_api = "responses"env_key = "ISH_API_KEY"
Codex’s custom-provider docs support this shape: provider ID, base URL, wire API, and environment-backed auth key.
Export your key
In the same shell where you start Codex:
export ISH_API_KEY="ish_live_your_key_here"
codex
For fish:
set -gx ISH_API_KEY "ish_live_your_key_here"
codex
Verify the setup
Start with a tiny text-only request:
codex exec --skip-git-repo-check "Reply exactly: ISH_OK"
Then verify the tool loop:
printf "TOOL_OK\n" > hello.txt
codex exec --skip-git-repo-check "Read hello.txt and reply exactly with its contents."
If both work, Codex can reach ISH, receive model output, and complete a local tool-read loop.
Supported models
Good defaults:
| Codex model | Use for |
|---|---|
claude-haiku-4.5 | quick checks and small edits |
claude-sonnet-4.6 | normal coding work |
claude-opus-4.8 | heavier reasoning and planning |
claude-sonnet-5 | newer Sonnet-class work when available |
gpt-5.6-sol | GPT-style reasoning when available |
Use canonical model IDs in Codex CLI. Cursor-style ish-... aliases are accepted by ISH, but canonical IDs keep CLI configs easier to read.
Track usage
Every Codex request appears in api.ish.chat/usage. You can see:
- key name and prefix
- endpoint type
- model
- input and output usage
- charged amount
- status
This is especially helpful for Codex because tool calls and project context can make a request larger than the message you typed.
Common errors
Empty response
Make sure the ISH provider has:
wire_api = "responses"
Without the Responses wire format, Codex may not speak the expected protocol for tool-call loops.
401 authentication error
Check that ISH_API_KEY is set in the same shell where Codex starts, and that the value starts with ish_live_.
Model not supported
Use a canonical Rootnull model ID such as:
claude-sonnet-4.6
Then restart Codex so the provider and model selection reload cleanly.
Insufficient balance
Add balance from api.ish.chat/dashboard. Codex traffic is API usage, so it does not spend free daily web chat usage.
Best default
Start with claude-sonnet-4.6. It is the safest balance for normal coding tasks. Move to claude-haiku-4.5 for cheap quick checks, and only move up to Opus-class models when the task needs it.
Useful links:
