OpenAI Codex + NeoSmith

Codex (OpenAI’s CLI coding agent) speaks the OpenAI Responses API (/v1/responses). NeoSmith implements this surface, so Codex can route through NeoSmith with a base-URL + key change.

• Endpoint: https://router.neosmith.ai/v1
• Format: OpenAI Responses API (/v1/responses)
• Model: neosmith.intelligent-pro

---

Setup

Codex reads OpenAI environment variables. Set them in your shell profile (see your platform guide):

export OPENAI_BASE_URL=https://router.neosmith.ai/v1
export OPENAI_API_KEY=sk-plus-yourname-xxxxxx

Then point Codex at the NeoSmith model. In Codex’s config (~/.codex/config.toml or equivalent), set the model and base URL:

model = "neosmith.intelligent-pro"
model_provider = "neosmith"

[model_providers.neosmith]
name = "NeoSmith"
base_url = "https://router.neosmith.ai/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"

The exact config key names depend on your Codex version. The essentials are: base URL https://router.neosmith.ai/v1, API key = your NeoSmith key, wire format = responses, model = neosmith.intelligent-pro.

---

What NeoSmith supports for Codex

NeoSmith’s /v1/responses implementation covers the surface Codex’s agentic loop needs:

Codex requirement	Supported
POST /v1/responses endpoint	✅
SSE streaming with Responses event types	✅
Function/tool calls in Responses format	✅
id + status fields on output items (multi-turn replay)	✅
store: true parameter	✅ (accepted; router is stateless)
reasoning.effort passthrough	✅ (routed/ignored per backend)
previous_response_id (conversation continuity)	✅ (accepted)

---

Verify

curl -s https://router.neosmith.ai/v1/responses \
  -H "Authorization: Bearer sk-plus-yourname-xxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "neosmith.intelligent-pro",
    "input": "Write a haiku about caching.",
    "max_output_tokens": 60
  }'

Expected: a Responses-API body with an output array of items, plus neosmith_meta.

Notes & limits

• Codex’s built-in tools (apply_patch, shell) are passed through in Responses format; NeoSmith accepts function-type tools. Non-function built-in tool types are accepted without error but may be handled by Codex client-side.
• File-input items and server-side storage are accepted but not persisted (the router is stateless by design — conversation state is reconstructed from the input each turn).

Troubleshooting

See Troubleshooting. If Codex errors on model validation, confirm you set model = "neosmith.intelligent-pro" (not gpt-*, which NeoSmith rejects with 400).