Codex CLI

Codex is the one client on this list that cannot be wired with environment variables, and the reason is worth understanding before you start: Codex ignores OPENAI_API_KEY and OPENAI_BASE_URL outright, and since February 2026 it speaks only the OpenAI Responses API — wire_api = "chat" is gone, and a config that still sets it now refuses to load. Any instructions you find that export two variables and run codex predate that change and will quietly send your traffic to OpenAI instead.

UsePod serves /v1/responses, so the correct wiring is a provider block.

Setup

Create or edit ~/.codex/config.toml:

model = "deepseek-v4-1-flash"
model_provider = "usepod"

[model_providers.usepod]
name = "UsePod"
base_url = "https://api.usepod.ai/proxy/<TOKEN>/v1"
wire_api = "responses"

Then run codex. Replace <TOKEN> with the API token from your dashboard.

No API key is required anywhere in this file. With no env_key set, Codex sends no Authorization header at all, and the token in the base URL is what authenticates the request — so there is no secret to export and nothing to leak into your shell history.

Set wire_api explicitly even though the documentation calls it the default; configs that omitted it have been observed failing to reach the provider.

Verify

Send a short prompt in the Codex TUI and watch the reply stream. Or confirm the endpoint independently first:

curl -s https://api.usepod.ai/proxy/<TOKEN>/v1/responses \
  -H 'content-type: application/json' \
  -d '{"model":"deepseek-v4-1-flash","input":"Reply with exactly: ok",
       "max_output_tokens":200}'

A "status": "completed" response with an output item containing ok means the surface is live for your token.

Choosing a model

Any model id from the marketplace works in the model field. One consequence is worth knowing, because it is invisible until you hit it: Codex decides which tools to offer the model based on the model id, matched against its own table of OpenAI names. Point it at an id it does not recognise — which includes every non-OpenAI model — and two things change.

It prints a startup warning:

warning: Model metadata for `deepseek-v4-1-flash` not found.
Defaulting to fallback metadata; this can degrade performance and cause issues.

That warning is about Codex’s local assumptions, not about UsePod rejecting anything; your requests route normally. It means Codex is guessing at the context window rather than reading it from a table, so it may compact the conversation earlier than necessary.

And it omits the apply_patch tool, offering a sub-agent namespace instead. Codex still edits files — it falls back to shell commands through exec_command — but multi-file edits go through a less structured path than they would on a model it recognises.

Neither behaviour breaks the integration. If you would rather have the patch tool, name a model Codex knows; if you would rather have the price, name the model you want and accept the warning.

What the Responses surface supports

• Streaming and tool calls both work, including multi-turn tool loops.
• Reasoning is passed through: models that emit reasoning content surface it as a reasoning item alongside the message.
• It is stateless. previous_response_id is rejected with a 400, because UsePod stores no responses. Codex sends store: false and resends the full conversation each turn, so this never comes up in normal use.
• Hosted tools are ignored. web_search and tool_search have no equivalent on a chat-completions upstream and are dropped rather than forwarded, which is what keeps upstreams from rejecting the whole request.

Troubleshooting

wire_api = "chat" is no longer supported. Delete the line or set it to "responses". Chat-completions support was removed from Codex in early 2026.

Requests go to api.openai.com. Your provider block is not being used — usually a missing model_provider = "usepod" at the top level, or a [model_providers.…] table whose name does not match it.

404 on /v1/responses. The base URL is missing its /v1 suffix. Unlike the Anthropic surface, this one takes it.

The startup metadata warning. Expected on any non-OpenAI model id. See above; it is not an error.