LM Warden
Language: English
Documentation Download & Install

Codex CLI on your own GPUs.

LM Warden as a custom model provider for Codex CLI, one served model.

Where a request from Codex CLI goes Codex CLI sends its requests to LM Warden with base_url set to the warden and env_key holding a warden key. LM Warden checks the key, speaks OpenAI Responses, queues the request by key priority and hands it to the replica router, which keeps each session on the replica that already holds its cache, one model copy per GPU. Codex CLI base_url the warden, at /v1 env_key LMWARDEN_KEY holds vw_… POST /v1/responses LM Warden checks the key: one per app speaks OpenAI Responses scheduler, key priority usage, latency and live stats per key Replica router a session goes back to the replica that holds its cache one model copy per GPU your GPUs: vLLM or llama.cpp replica 0, GPU 0 replica 1, GPU 1 replica 2, GPU 2 a conversation stays on its replica
Where a request from Codex CLI goes Codex CLI sends its requests to LM Warden with base_url set to the warden and env_key holding a warden key. LM Warden checks the key, speaks OpenAI Responses, queues the request by key priority and hands it to the replica router, which keeps each session on the replica that already holds its cache, one model copy per GPU. Codex CLI base_url the warden, at /v1 env_key LMWARDEN_KEY holds vw_… POST /v1/responses LM Warden checks the key: one per app speaks OpenAI Responses scheduler, key priority usage, latency, live stats per key Replica router a session goes back to the replica that holds its cache replica 0 GPU 0 replica 1 GPU 1 replica 2 GPU 2 your GPUs: vLLM or llama.cpp a conversation stays on its replica
Where a request from Codex CLI goes: the two settings Connect writes, the warden in the middle, then one copy of the model per GPU.

In words: Codex CLI sends its requests to LM Warden with base_url set to the warden and env_key holding a warden key. LM Warden checks the key, speaks OpenAI Responses, queues the request by key priority and hands it to the replica router, which keeps each session on the replica that already holds its cache, one model copy per GPU.

Protocol
OpenAI Responses
Support
Local models only

Local models only. Codex speaks only the Responses API; the warden translates POST /v1/responses to Chat Completions for the served model, so the reply, tool calls and usage reach Codex as Responses events. It is stateless: Codex sends store:false and the whole conversation each turn, and previous_response_id is refused. The thinking is on by default (Codex asks for reasoning summaries and the model’s reasoning is shown as a summary; to turn it off, set model_reasoning_effort = "none" in config.toml), hosted web search is dropped (set web_search = "disabled"), and router rules do not apply. Set model_context_window to the model’s real window: Codex assumes 272k for names it does not know. The engine needs tool calling enabled (vLLM: --enable-auto-tool-choice and a --tool-call-parser). A request for a model that is not loaded gets 404 and Codex retries for several seconds before it gives up.

Requirements

  • Needs a model with tool calling (agentic edits).
  • Works best with at least 32,768 tokens of context.

Setup

In the console, Connect writes these files with your warden’s address, your key and the model you picked. Here they are with placeholders.

Local models

Add LM Warden as a model provider in ~/.codex/config.toml.

config.toml (~/.codex/config.toml)
model = "your-served-model-name"
model_provider = "lmwarden"
# Codex assumes a 272k window for models it does not know; set the real one.
model_context_window = 32768

[model_providers.lmwarden]
name = "LM Warden"
base_url = "https://your-warden/v1"
env_key = "LMWARDEN_KEY"
wire_api = "responses"
# The model's context window is unknown: 32768 is a guess. Set it to the model's real window.
Shell
export LMWARDEN_KEY=vw_YOUR_KEY
codex exec "Reply with the single word OK."

Official documentation: https://learn.chatgpt.com/docs/config-file/config-reference (read 2026-10-04).

Related: all coding agents · let your agent pick the model · API keys