Connect Codex to third-party models through Responses

The short answer

Current Codex supports only "responses" for wire_api. "chat" was removed and causes a startup configuration error. KimiSeek provides /v1/responses, so the configuration below connects directly.

30-second decision table

Why Codex defaults to responses

Codex sends requests to /responses and consumes its tool-call and streaming events. Chat Completions uses different request and event structures; changing wire_api cannot make current Codex use it. Model capabilities also depend on the upstream and gateway implementation. A successful endpoint probe does not validate every tool capability.

Prerequisites

• Node.js 20 or newer
• Codex CLI: npm install -g @openai/codex
• A KimiSeek API key (starts with ks-), created in the dashboard

Step 1 · Write the config file

Edit ~/.codex/config.toml (create it if it does not exist):

model = "deepseek-v4-pro-0813"
model_provider = "kimiseek"

[model_providers.kimiseek]
name = "KimiSeek"
base_url = "https://api.kimiseek.app/v1"
env_key = "KIMISEEK_API_KEY"
wire_api = "responses"

Three things that trip people up:

1. base_url must include /v1. Codex appends /responses directly to base_url. Drop the /v1 and the request goes to https://api.kimiseek.app/responses — a 404.
2. env_key is a variable name, not the key itself. You write KIMISEEK_API_KEY here; the actual value lives in your environment.
3. model must be an id that really exists in the catalog. A typo returns model not found, and so does a retired model id.

Step 2 · Set the environment variable and start

export KIMISEEK_API_KEY="ks-your-key"

codex

To switch models temporarily, edit model in config.toml or pass codex --model <id>.

Step 3 · Confirm the key works

curl https://api.kimiseek.app/v1/models \
  -H "Authorization: Bearer ${KIMISEEK_API_KEY:?Set KIMISEEK_API_KEY before running this command}"

If you get a model list back, both the key and the network are fine. The point of this step is to separate authentication failures from protocol failures — otherwise a 404 leaves you guessing which one it was.

Step 4 · Probe responses in 30 seconds

curl https://api.kimiseek.app/v1/responses \
  -H "Authorization: Bearer ${KIMISEEK_API_KEY:?Set KIMISEEK_API_KEY before running this command}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro-0813",
    "input": "Explain what a base URL is in one sentence.",
    "max_output_tokens": 256,
    "stream": false
  }'

• HTTP 200 with a JSON output array → the Responses route works; next test a real task and tool calls.
• HTTP 404 → check that base_url includes /v1, then inspect whether the response reports a missing route or a missing model. The status alone is not enough.
• HTTP 401 → return to Step 3 and check credentials.

If the gateway only offers Chat Completions

Keep wire_api = "responses". Choose a gateway providing /v1/responses, or add a compatible translation layer. Do not switch to the removed "chat" value.

Common errors

More protocol-level errors are collected in the error reference.

Cost

Codex usage is input-heavy: every turn re-sends the context. DeepSeek V4 Pro is priced in CNY per million tokens across two bands:

Peak is exactly double. Batching large refactors into off-peak hours halves the bill. Live numbers for every model come from the pricing page, which reads the model catalog directly.

Versions and verification

Configuration checked with Codex CLI 0.154.0 on 2026-10-26. See the official configuration reference for protocol support.