This is the machine’s HTTP.
Live on this host. Same origin, same routes the desk calls.
A platform Bearer or the panel cookie opens the gate —
there is no OpenAI sk- field. Usage bills the shared ChatGPT plan.
- must Required. Path parameters always. Body, query, or header when OpenAPI
requiredis true. - should Recommended. A default exists, or you must pick one of a pair (
textorimages;cronorrun_at). - optional Skip unless you need the behavior.
Connect IDE
Cursor or Roo Code
Mint a per-user bg_… key, use this host's
OpenAI-compatible base URL, then choose a model from
GET /v1/models.
- Base URL
/v1- Models
gpt-5.6-sol(default) ·gpt-6-astra- Auth
Authorization: Bearer bg_…
curl -N -H "Authorization: Bearer $CODEX_GATEWAY_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"gpt-5.6-sol","reasoning_effort":"low","stream":true,"messages":[{"role":"user","content":"Say IDE_OK"}]}' \
/v1/chat/completions
Chat/Ask is supported. Client-defined IDE tool calls are not; use a
platform coding chat/job for shell and repository edits. Full setup:
docs/38-cursor-and-roo-code.md.