Integrate LMDeploy with Codex

Codex can connect to an OpenAI Responses-compatible gateway by configuring a custom model provider. LMDeploy exposes POST /v1/responses, so Codex can route requests to a local or remote LMDeploy api_server.

The request path is:

Codex -> http://<server>:<port>/v1/responses -> LMDeploy api_server

1. Start LMDeploy

Launch an LMDeploy API server with the model you want Codex to use:

lmdeploy serve api_server Qwen/Qwen3.5-35B-A3B --backend pytorch --server-port 23333

For tool calling, launch the server with a tool parser supported by your model:

lmdeploy serve api_server <model> \
  --server-port 23333 \
  --tool-call-parser <parser-name>

2. Verify the Responses Endpoint

Before configuring Codex, check that LMDeploy responds to Responses-style requests:

curl http://127.0.0.1:23333/v1/responses \
  -H "content-type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.5-35B-A3B",
    "input": "Say hello from LMDeploy",
    "max_output_tokens": 32
  }'

The model value must match a model name exposed by your LMDeploy server. You can check the model list with:

curl http://127.0.0.1:23333/v1/models

3. Configure Codex

Add the LMDeploy provider configuration to ~/.codex/config.toml:

model = "Qwen/Qwen3.5-35B-A3B"
model_provider = "lmdeploy"

[model_providers.lmdeploy]
name = "LMDeploy"
base_url = "http://127.0.0.1:23333/v1"
env_key = "LMDEPLOY_API_KEY"
wire_api = "responses"
requires_openai_auth = false
stream_idle_timeout_ms = 300000

Set base_url to the /v1 root, not to /v1/responses. Codex appends /responses itself when wire_api = "responses".

Set LMDEPLOY_API_KEY before starting Codex. If the LMDeploy server was not launched with --api-keys, any non-empty value is sufficient. Otherwise, use the matching key instead of dummy:

export LMDEPLOY_API_KEY=dummy

4. Run Codex

Start Codex:

codex

Or run a single non-interactive request:

codex exec "Say hello from LMDeploy"

5. Tool Calling

Codex commonly uses tool calls for shell commands and file inspection. LMDeploy's /v1/responses endpoint supports function-call output when the API server is launched with --tool-call-parser.

6. Endpoint Note

Codex uses the OpenAI Responses-compatible /v1/responses endpoint. The Anthropic-compatible /v1/messages endpoint is for Anthropic-style clients such as Claude Code.