Codex → AI24X

1) Prerequisites

Sign up → Console → create an API key → keep a small credit balance. Codex CLI ≥ 0.146 (Responses wire API) or desktop. Paste the key in full—never truncate with "…". VIP named models need an active membership.

2) Provider config (~/.codex/config.toml)

CLI and desktop share ~/.codex/config.toml. Add the ai24x provider: base_url must end at https://api.ai24x.com/v1, and wire_api must be "responses" (Codex 0.146 deprecated chat):

Restart the desktop app after saving; open a new terminal for CLI. Start with flash or vip-ds-flash—see the model table in section 5.

3) Set the API key (env var)

Keep the key in an environment variable only—never paste it into a Codex chat (it triggers a secret-leak warning and can stall the session):

setx does not affect already-open windows—open a new terminal, or set $env:AI24X_API_KEY="…" temporarily to verify in the current window.

4) Smoke test

Non-interactive check (expect OK, no errors):

Or verify /v1/responses outside the IDE with curl (live since 2026-08-05, text chat returns 200):

Expect HTTP 200, object: "response", and text in output[0].content.

5) DeepSeek & VIP named models

Switch models by changing model = to an id below. VIP ids (vip-*) need an active membership and spend prepaid credits; higher multipliers drain balance faster. Prices are point-in-time estimates—billing follows the Console.

Full catalog of 20+ named models (Kimi / Qwen / GLM / MiniMax / GPT-5 / Claude / Gemini …): VIP list.

5.1) Full VIP alias catalog (24)

All 24 vip-* aliases (live on the gateway since 2026-08-11). Use them with -c "model=vip-xxx" (section 6). New additions: GPT-5 / GPT-5.4 / GPT-5 mini.

6) Pick a model & switch in-session (/model)

CLI (recommended): pass the provider and model on every run so the alias always routes to AI24X:

Inside an interactive session, type /model and pick any alias from the catalog above — switch anytime without starting a new session or losing context.

Pitfalls

• -m "provider/model" combined form errors out — e.g. -m "ai24x/vip-gpt5" is invalid. Do not join provider and model with a slash.
• -m "vip-xxx" alone may route to the wrong provider — always pair -c "model_provider=…" with -c "model=…".
• Desktop model picker does not list AI24X models — upstream gating, not planned. Use /model or CLI flags; do not expect the picker to show them.
• Replace ai24x-prod with the provider name defined in your ~/.codex/config.toml (the snippet in section 2 names it ai24x).

7) Status & pitfalls

• Responses is live; agent tool loops are being completed — AI24X /v1/responses handles text chat (verified 200); full agent tool calling runs on /v1/chat/completions (OpenAI tools supported). Codex currently talks Responses: text and planning work; auto-editing tool loops unlock once the platform completes Responses tool support.
• base_url missing /v1 — requests hit wrong paths like /chat/completions. Use https://api.ai24x.com/v1.
• 401 — bad, truncated, or revoked key. Paste the full key from Console; create a new one if unsure.
• 402 — insufficient balance / prepaid needed for the model. Top up in Console.
• 403 named model needs active membership — vip-* requires membership + prepaid.
• "Model metadata not found" warning is harmless — custom providers fall back to default metadata; it does not affect usage.
• Don't combine -m "provider/model" — errors out; use -c "model_provider=…" + -c "model=…" (section 6).
• Desktop picker won't show AI24X models — upstream gating; switch with /model or CLI flags.

Create account All integrations OpenClaw guide VIP model list Console