@reasoning-router/proxy

English | 简体中文 | 日本語 | 한국어

A standalone HTTP proxy that asks a classifier how much reasoning each request needs, then applies that effort to the outgoing Responses (POST /v1/responses) or Anthropic Messages (POST /v1/messages) request without breaking the prompt cache. Requires Node.js 24.x.

Alpha. For trusted local use only; read the trust model first.

Install

npm install -g @reasoning-router/proxy   # installs the reasoning-router command

Or run it without installing: npx @reasoning-router/proxy.

Usage

Create a .env in the directory you run from (or export the variables):

REASONING_ROUTER_CLASSIFIER_API_KEY=your-jev-key
# REASONING_ROUTER_CLASSIFIER_BASE_URL=https://ai-gateway.vercel.sh/typesafe   # Vercel keys only
REASONING_ROUTER_UPSTREAM_BASE_URL=https://api.openai.com/v1
REASONING_ROUTER_UPSTREAM_AUTH=bearer
REASONING_ROUTER_UPSTREAM_API_KEY=your-endpoint-key

Then start it and point your client at http://127.0.0.1:4320/v1:

reasoning-router
curl --fail http://127.0.0.1:4320/ready

Run reasoning-router --help for every variable. See .env.example for defaults, docs/environment.md for the classifier settings, and docs/behavior.md for probes, shutdown, limits, and forwarding.

Trust model

The proxy is for a single trusted machine. It listens only on 127.0.0.1 and has no caller authentication: any local process that can reach the port can send requests, and with REASONING_ROUTER_UPSTREAM_AUTH=bearer those requests use your configured upstream key.

Every request must carry a Host of 127.0.0.1:<port> or localhost:<port> for the port the proxy listens on; anything else gets a 400 before the body is read, classified, or forwarded. This blocks DNS-rebinding attacks from web pages. It is not authentication, and it does not stop a web page from sending requests to http://127.0.0.1:<port> directly. Do not expose the port through a tunnel or reverse proxy; a forward or container mapping to a different port also gets 400 invalid_host.

Codex CLI

Start the proxy as above, then create a Codex profile at ~/.codex/reasoning-router.config.toml. It keeps the model name and changes only the provider, so your default Codex setup is untouched:

model = "gpt-6-astra"
model_provider = "reasoning-router"

[model_providers.reasoning-router]
name = "reasoning-router"
base_url = "http://127.0.0.1:4320/v1"
wire_api = "responses"

Run codex -p reasoning-router to route through the proxy; plain codex is unchanged. -p names the profile file, not the provider; both are called reasoning-router here. The profile is not tied to one model: pick another with -m, for example codex -p reasoning-router -m gpt-6-sol. Profiles (-p) need Codex 0.161.0 or later. On older versions, put the same lines in ~/.codex/config.toml instead.

Codex sends no credential; the proxy adds REASONING_ROUTER_UPSTREAM_API_KEY (REASONING_ROUTER_UPSTREAM_AUTH=bearer). Leave supports_websockets unset, since the proxy serves HTTP only.

• Keep the real model ID. Codex reads model metadata by name, and the proxy has no aliases, so a made-up name like astra-rr would get generic metadata from Codex and a local 400 from the proxy.
• Use gpt-6-astra, gpt-6-luna, gpt-6-sol, or gpt-6.1-sol. Other models get a local 400.
• The classifier picks the effort. Codex's model_reasoning_effort and /model effort are ignored, though Codex still displays them.
• The proxy closes an upstream stream after 60 seconds without data. Raise REASONING_ROUTER_UPSTREAM_IDLE_TIMEOUT_MS if long turns are cut off.

ChatGPT subscription through CLIProxyAPI

The proxy cannot use a ChatGPT login directly. To use a subscription, run CLIProxyAPI with a Codex login and put it behind the proxy:

codex -p reasoning-router → reasoning-router (127.0.0.1:4320) → CLIProxyAPI (127.0.0.1:8317) → ChatGPT

Start the proxy in forward mode against CLIProxyAPI, with no upstream key:

REASONING_ROUTER_UPSTREAM_BASE_URL=http://127.0.0.1:8317/v1
REASONING_ROUTER_UPSTREAM_AUTH=forward

Then add env_key to the profile's provider block so Codex sends your CLIProxyAPI client key, which the proxy forwards unchanged:

env_key = "CLIPROXY_KEY"   # environment variable holding a CLIProxyAPI api-key

CLIProxyAPI passes configuration_update through for GPT-6 models on Codex Plus, Pro, and Team plans; on the Free plan it strips it for every model except gpt-6-luna, so routing has no effect there. Verified with Codex 0.161.0 and CLIProxyAPI 8.0.23: the backend accepted effort changes and kept the prompt cache across them.

See docs/harnesses/codex.md for details.

License

MIT