Skip to main content
FireConnect routes the OpenAI Codex CLI through Fireworks AI models via the Responses API. See the FireConnect overview for install and CLI basics.
Change models: fireconnect codex on --model <id>. See Models.

​
Prerequisites

Fire Pass keys (fpk_...) are not supported for Codex yet. The /responses endpoint requires a standard Fireworks API key.

​
Enable Fireworks routing

fireconnect codex and fireconnect chatgpt share one config — both the Codex CLI and the ChatGPT desktop app route through Fireworks with a single command.
Quit the ChatGPT app before on or off so its model list refreshes. Same quit rules as Cursor and VS Code.
Or pass the key once:
After fireconnect codex on or off, config.toml is updated immediately. To use updated routing, exit Codex and resume with codex resume <id>, or start a new session.

​
Default model

Codex routes a single default model. The default is kimi-fast-latest.

​
What gets written

FireConnect edits ~/.codex/config.toml:
  • Sets root model_provider / model for Codex 0.134+ (stored as a short slug)
  • Adds a [model_providers.fireworks-ai] block with wire_api = "responses" and a baked experimental_bearer_token literal (file mode 0600). Codex reads the key from config; no shell hook is required.
FireConnect snapshots your original ~/.codex/config.toml before the first change. The snapshot lives in ~/.fireconnect/codex/. Running fireconnect codex off restores it byte-for-byte. Unrelated Codex settings (for example [[mcp_servers]]) are preserved via surgical TOML edits.

​
Codex model catalog

When you run fireconnect codex on, FireConnect fetches your account’s serverless catalog and writes Codex-compatible model metadata to ~/.codex/fireworks-model-catalog.json. It links that file from config.toml via model_catalog_json so Codex knows display names, context windows, reasoning levels, and tool-calling support for each model.
The catalog includes serverless models that support tool calling with a non-zero context window, plus curated aliases such as glm-fast-latest. Deprecated models are excluded. If catalog generation fails (for example, due to an invalid API key), routing still works but Codex may show limited model metadata until you re-run fireconnect codex on. Browse available models globally:

​
FireRouter

Route requests through FireRouter:
FireConnect rejects --model firerouter with Fire Pass (fpk_...) on every harness; use an fw_... key. Codex also does not support Fire Pass for direct routing. FireRouter BYOK reads ANTHROPIC_API_KEY from your shell (Codex does not support --routing-preference).
MiniMax models are not supported in Codex. Codex uses the Fireworks Responses API and may insert assistant messages between tool_calls and tool_results. MiniMax chat templates require tool_results to follow tool_calls directly. Use Chat Completions harnesses (for example Claude Code or OpenCode) for MiniMax.

​
CLI reference

Run fireconnect codex help for all options.

​
Switch models

Some models expose multiple reasoning levels in the Codex catalog (for example, glm-5p2 supports high and max). Pick the model in Codex with /model after switching.

​
Turn off Fireworks routing

This restores your previous config.toml from the backup in ~/.fireconnect/codex/.

​
Use a non-default config file

​
Fireworks on Microsoft Foundry

Codex supports Fireworks on Microsoft Foundry (CLI: --provider azure or on --azure). See the FireConnect overview and Microsoft Foundry integration guide for portal setup.
Foundry routing requires a standard Azure API key. Fire Pass keys (fpk_...) are not supported. FireConnect does not write a Fireworks model catalog on the Foundry path. Set your deployment with --model.

​
Configure and enable

One-off routing:

​
What gets written

FireConnect sets model_provider = "fireworks-azure" in config.toml with a Fireworks on Microsoft Foundry provider block pointed at your Foundry endpoint. The Azure API key is baked as a literal when passed with --api-key, or referenced via env_key = "AZURE_API_KEY" when resolved from the environment. Pass your Foundry model with --model (for example, FW-GLM-5.2). Use fireconnect model list only for browsing Fireworks serverless models on the direct gateway path.

​
Turn off Foundry routing

To switch back to the Fireworks gateway:
To remove FireConnect entirely and restore your original config.toml:
See Turn off Foundry routing for details on global config behavior and uninstall.

​
Source

FireConnect is open source: github.com/fw-ai/fireconnect