Codex (local)

Run OpenAI's Codex agent on your machine — no API key, using your ChatGPT login.

The codex provider runs OpenAI's Codex agent locally instead of calling a plain model. Scira's research loop, tools, and instructions run inside the Codex harness on your real files, authenticated by your existing codex login session — no API key, using your ChatGPT plan.

Set up

Install Codex

npm install -g @openai/codex

See the Codex docs for other install options, including Homebrew.

Log in

codex login

Authentication uses your ChatGPT account session (stored under ~/.codex, e.g. ~/.codex/auth.json). Scira reads only the logged-in account, never an API key.

Select it in Scira

From the TUI, run /llm → Codex (local). Or set it in config:

~/.scira/config.json

{ "llmProvider": "codex", "model": "gpt-5.5" }

Available models

Model ID	Description
`gpt-5.5`	Default. Most capable Codex model
`gpt-5.4`	Previous flagship
`gpt-5.4-mini`	Smaller, faster, lower cost
`gpt-5.3-codex-spark`	Lightweight Codex variant

Pick one with /model.

Reasoning effort

For reasoning-capable models, set effort with /reasoning, mapped to harness.reasoningEffort in config:

Level	Effect
`low`	Fastest, least deliberation
`medium`	Default. Balanced
`high`	Most deliberation on harder turns

~/.scira/config.json

{
  "llmProvider": "codex",
  "harness": { "reasoningEffort": "medium" }
}

Notes

Web search runs through Scira's own Exa/Firecrawl/Parallel pipeline, so answers cite your configured sources.
Runs use your local filesystem and process access in the current project directory — there's no cloud sandbox.
scira doctor reports whether codex is on your PATH and whether you're logged in (read from ~/.codex/auth.json).

No API key

Scira strips credential environment variables before launching the harness, so OPENAI_API_KEY is ignored for this provider — your ChatGPT login is used instead.